Training in progress, step 50, checkpoint

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-50/tokenizer.json filter=lfs diff=lfs merge=lfs -text

checkpoint-50/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "sliding_window": 4096,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": false,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-50/generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151646,
+  "do_sample": true,
+  "eos_token_id": 151643,
+  "temperature": 0.6,
+  "top_p": 0.95,
+  "transformers_version": "4.51.3"
+}

checkpoint-50/global_step50/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04d7c4a753ad8773e2bc43bcfeb2976f8837949927125a64e5587cfc2187b89e
+size 5331274140

checkpoint-50/global_step50/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd6cd1723ab6752efaea5a922408fdf67cfef7c8af98509f7d8384e09b992cb6
+size 5331276572

checkpoint-50/global_step50/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2523a1882f16bcf66cf73a7b1d0710fba96bfe430068dddc0a8682ae23c6686
+size 5331276892

checkpoint-50/global_step50/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70e5c0eda53efef25069fb4e290ad2de62b0a6bb07153b0c2803465ac175808d
+size 5331273884

checkpoint-50/global_step50/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3168d276e15e44440db87bd9594897cd34871b3a5ca24040924d51bbb4b47d0
+size 3554267640

checkpoint-50/latest

@@ -0,0 +1 @@
+global_step50

checkpoint-50/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:396f0d1abcb130e9903eccdcabcb61e7502a8ded1258d55c9749c124dee7c71a
+size 3554214752

checkpoint-50/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2702d282c4493893603dc2a13af95d09e111a64c63f52fb9af5255406ae8d654
+size 14960

checkpoint-50/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:430e93858450810d232345690d663543e8bbf2a60c00a022392222fb16ec625f
+size 14960

checkpoint-50/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:158ffd4abe490709b9e9d3b062da97147f2d078f352a84f7c20b893ff17f6452
+size 14960

checkpoint-50/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67b5fbbcde50895969936bc2537f01f041594a2ca632fbec18fb08928d76422d
+size 14960

checkpoint-50/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0208179da2b605778b21720a99ccfb3d5e515115ee90824c90bfcabf8ad99120
+size 1064

checkpoint-50/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-50/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4256422650d141f228fe954acee98679da412984c29a569877eefd3af69315a
+size 11422959

checkpoint-50/tokenizer_config.json

@@ -0,0 +1,195 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<｜end▁of▁sentence｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<｜User｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151645": {
+      "content": "<｜Assistant｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151646": {
+      "content": "<｜begin▁of▁sentence｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|EOT|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151648": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151649": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<｜begin▁of▁sentence｜>",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin��>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<｜end▁of▁sentence｜>",
+  "extra_special_tokens": {},
+  "legacy": true,
+  "model_max_length": 16384,
+  "pad_token": "<｜end▁of▁sentence｜>",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizerFast",
+  "unk_token": null,
+  "use_default_system_prompt": false
+}

checkpoint-50/trainer_state.json

@@ -0,0 +1,1384 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.05714285714285714,
+  "eval_steps": 500,
+  "global_step": 50,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1734.0,
+      "completions/mean_length": 1702.03125,
+      "completions/mean_terminated_length": 993.6190795898438,
+      "completions/min_length": 483.0,
+      "completions/min_terminated_length": 483.0,
+      "epoch": 0.001142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20054349303245544,
+      "learning_rate": 0.0,
+      "loss": 0.0427,
+      "num_tokens": 118418.0,
+      "reward": 0.17899775505065918,
+      "reward_std": 0.7650213241577148,
+      "rewards/cosine_scaled_reward/mean": -0.09800112992525101,
+      "rewards/cosine_scaled_reward/std": 0.37953105568885803,
+      "rewards/format_reward/mean": 0.375,
+      "rewards/format_reward/std": 0.48795005679130554,
+      "step": 1
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1894.0,
+      "completions/mean_length": 1738.90625,
+      "completions/mean_terminated_length": 949.0,
+      "completions/min_length": 435.0,
+      "completions/min_terminated_length": 435.0,
+      "epoch": 0.002285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.19502359628677368,
+      "learning_rate": 5e-08,
+      "loss": 0.0561,
+      "num_tokens": 239748.0,
+      "reward": 0.3848632574081421,
+      "reward_std": 0.9111153483390808,
+      "rewards/cosine_scaled_reward/mean": 0.020556632429361343,
+      "rewards/cosine_scaled_reward/std": 0.4492928683757782,
+      "rewards/format_reward/mean": 0.34375,
+      "rewards/format_reward/std": 0.4787135720252991,
+      "step": 2
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.90625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1287.0,
+      "completions/mean_length": 1944.453125,
+      "completions/mean_terminated_length": 943.5,
+      "completions/min_length": 608.0,
+      "completions/min_terminated_length": 608.0,
+      "epoch": 0.0034285714285714284,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.230765700340271,
+      "learning_rate": 1e-07,
+      "loss": 0.0549,
+      "num_tokens": 374665.0,
+      "reward": -0.28856638073921204,
+      "reward_std": 0.4003669023513794,
+      "rewards/cosine_scaled_reward/mean": -0.19897069036960602,
+      "rewards/cosine_scaled_reward/std": 0.18252794444561005,
+      "rewards/format_reward/mean": 0.109375,
+      "rewards/format_reward/std": 0.3145764470100403,
+      "step": 3
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1871.0,
+      "completions/mean_length": 1592.3125,
+      "completions/mean_terminated_length": 1006.4285888671875,
+      "completions/min_length": 450.0,
+      "completions/min_terminated_length": 450.0,
+      "epoch": 0.004571428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20995420217514038,
+      "learning_rate": 1.5e-07,
+      "loss": 0.1266,
+      "num_tokens": 486381.0,
+      "reward": 0.20640414953231812,
+      "reward_std": 0.8193759918212891,
+      "rewards/cosine_scaled_reward/mean": -0.13117292523384094,
+      "rewards/cosine_scaled_reward/std": 0.35454094409942627,
+      "rewards/format_reward/mean": 0.46875,
+      "rewards/format_reward/std": 0.5029674172401428,
+      "step": 4
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.953125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1680.0,
+      "completions/mean_length": 2002.859375,
+      "completions/mean_terminated_length": 1085.0,
+      "completions/min_length": 755.0,
+      "completions/min_terminated_length": 755.0,
+      "epoch": 0.005714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23816199600696564,
+      "learning_rate": 2e-07,
+      "loss": 0.01,
+      "num_tokens": 625380.0,
+      "reward": -0.41131818294525146,
+      "reward_std": 0.30660682916641235,
+      "rewards/cosine_scaled_reward/mean": -0.24472159147262573,
+      "rewards/cosine_scaled_reward/std": 0.19079075753688812,
+      "rewards/format_reward/mean": 0.078125,
+      "rewards/format_reward/std": 0.27048972249031067,
+      "step": 5
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1344.0,
+      "completions/mean_length": 1890.0,
+      "completions/mean_terminated_length": 784.0,
+      "completions/min_length": 440.0,
+      "completions/min_terminated_length": 440.0,
+      "epoch": 0.006857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24285951256752014,
+      "learning_rate": 2.5e-07,
+      "loss": -0.0119,
+      "num_tokens": 757988.0,
+      "reward": -0.24828195571899414,
+      "reward_std": 0.3839319050312042,
+      "rewards/cosine_scaled_reward/mean": -0.19445347785949707,
+      "rewards/cosine_scaled_reward/std": 0.19692479074001312,
+      "rewards/format_reward/mean": 0.140625,
+      "rewards/format_reward/std": 0.3503824472427368,
+      "step": 6
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1941.0,
+      "completions/mean_length": 1935.046875,
+      "completions/mean_terminated_length": 1390.8182373046875,
+      "completions/min_length": 886.0,
+      "completions/min_terminated_length": 886.0,
+      "epoch": 0.008,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2183438539505005,
+      "learning_rate": 3e-07,
+      "loss": 0.0412,
+      "num_tokens": 892239.0,
+      "reward": -0.07044821977615356,
+      "reward_std": 0.5991545915603638,
+      "rewards/cosine_scaled_reward/mean": -0.14459910988807678,
+      "rewards/cosine_scaled_reward/std": 0.3703240156173706,
+      "rewards/format_reward/mean": 0.21875,
+      "rewards/format_reward/std": 0.4166666865348816,
+      "step": 7
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1957.0,
+      "completions/mean_length": 1743.921875,
+      "completions/mean_terminated_length": 966.8333129882812,
+      "completions/min_length": 296.0,
+      "completions/min_terminated_length": 296.0,
+      "epoch": 0.009142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.18490855395793915,
+      "learning_rate": 3.5e-07,
+      "loss": 0.0096,
+      "num_tokens": 1014266.0,
+      "reward": 0.07391861081123352,
+      "reward_std": 0.5062483549118042,
+      "rewards/cosine_scaled_reward/mean": -0.11929068714380264,
+      "rewards/cosine_scaled_reward/std": 0.4095526933670044,
+      "rewards/format_reward/mean": 0.3125,
+      "rewards/format_reward/std": 0.467176616191864,
+      "step": 8
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1790.0,
+      "completions/mean_length": 1965.46875,
+      "completions/mean_terminated_length": 1461.111083984375,
+      "completions/min_length": 1029.0,
+      "completions/min_terminated_length": 1029.0,
+      "epoch": 0.010285714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21707069873809814,
+      "learning_rate": 4e-07,
+      "loss": 0.0566,
+      "num_tokens": 1151512.0,
+      "reward": -0.15350507199764252,
+      "reward_std": 0.7245944738388062,
+      "rewards/cosine_scaled_reward/mean": -0.18612754344940186,
+      "rewards/cosine_scaled_reward/std": 0.30883485078811646,
+      "rewards/format_reward/mean": 0.21875,
+      "rewards/format_reward/std": 0.4166666865348816,
+      "step": 9
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1745.0,
+      "completions/mean_length": 1682.59375,
+      "completions/mean_terminated_length": 817.1578979492188,
+      "completions/min_length": 394.0,
+      "completions/min_terminated_length": 394.0,
+      "epoch": 0.011428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20094214379787445,
+      "learning_rate": 4.5e-07,
+      "loss": 0.0457,
+      "num_tokens": 1270030.0,
+      "reward": 0.027805477380752563,
+      "reward_std": 0.4805509150028229,
+      "rewards/cosine_scaled_reward/mean": -0.14234726130962372,
+      "rewards/cosine_scaled_reward/std": 0.26565250754356384,
+      "rewards/format_reward/mean": 0.3125,
+      "rewards/format_reward/std": 0.467176616191864,
+      "step": 10
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.953125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1094.0,
+      "completions/mean_length": 1998.15625,
+      "completions/mean_terminated_length": 984.6666870117188,
+      "completions/min_length": 798.0,
+      "completions/min_terminated_length": 798.0,
+      "epoch": 0.012571428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2170705795288086,
+      "learning_rate": 5e-07,
+      "loss": 0.0247,
+      "num_tokens": 1409584.0,
+      "reward": -0.43332377076148987,
+      "reward_std": 0.36288702487945557,
+      "rewards/cosine_scaled_reward/mean": -0.24791188538074493,
+      "rewards/cosine_scaled_reward/std": 0.17533892393112183,
+      "rewards/format_reward/mean": 0.0625,
+      "rewards/format_reward/std": 0.24397502839565277,
+      "step": 11
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2044.0,
+      "completions/mean_length": 1630.375,
+      "completions/mean_terminated_length": 1093.4285888671875,
+      "completions/min_length": 427.0,
+      "completions/min_terminated_length": 427.0,
+      "epoch": 0.013714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2160935252904892,
+      "learning_rate": 5.5e-07,
+      "loss": 0.0753,
+      "num_tokens": 1524872.0,
+      "reward": 0.0067175328731536865,
+      "reward_std": 0.689138650894165,
+      "rewards/cosine_scaled_reward/mean": -0.22320374846458435,
+      "rewards/cosine_scaled_reward/std": 0.3645767867565155,
+      "rewards/format_reward/mean": 0.453125,
+      "rewards/format_reward/std": 0.501733124256134,
+      "step": 12
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1751.0,
+      "completions/mean_length": 1833.453125,
+      "completions/mean_terminated_length": 1067.21435546875,
+      "completions/min_length": 616.0,
+      "completions/min_terminated_length": 616.0,
+      "epoch": 0.014857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2122364640235901,
+      "learning_rate": 6e-07,
+      "loss": 0.0326,
+      "num_tokens": 1653253.0,
+      "reward": -0.09265299141407013,
+      "reward_std": 0.5985201001167297,
+      "rewards/cosine_scaled_reward/mean": -0.17913900315761566,
+      "rewards/cosine_scaled_reward/std": 0.306300550699234,
+      "rewards/format_reward/mean": 0.265625,
+      "rewards/format_reward/std": 0.44515693187713623,
+      "step": 13
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1897.0,
+      "completions/mean_length": 1823.40625,
+      "completions/mean_terminated_length": 1202.4705810546875,
+      "completions/min_length": 605.0,
+      "completions/min_terminated_length": 605.0,
+      "epoch": 0.016,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2076576203107834,
+      "learning_rate": 6.5e-07,
+      "loss": 0.0261,
+      "num_tokens": 1780559.0,
+      "reward": 0.005522748455405235,
+      "reward_std": 0.7086418867111206,
+      "rewards/cosine_scaled_reward/mean": -0.1378636360168457,
+      "rewards/cosine_scaled_reward/std": 0.35400503873825073,
+      "rewards/format_reward/mean": 0.28125,
+      "rewards/format_reward/std": 0.4531635046005249,
+      "step": 14
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1328.0,
+      "completions/mean_length": 1698.171875,
+      "completions/mean_terminated_length": 731.0,
+      "completions/min_length": 391.0,
+      "completions/min_terminated_length": 391.0,
+      "epoch": 0.017142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.1969502866268158,
+      "learning_rate": 7e-07,
+      "loss": 0.0216,
+      "num_tokens": 1900162.0,
+      "reward": 0.2789269685745239,
+      "reward_std": 0.43547046184539795,
+      "rewards/cosine_scaled_reward/mean": -0.00897398591041565,
+      "rewards/cosine_scaled_reward/std": 0.4515364170074463,
+      "rewards/format_reward/mean": 0.296875,
+      "rewards/format_reward/std": 0.4604927599430084,
+      "step": 15
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 1.0,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 0.0,
+      "completions/mean_length": 2048.0,
+      "completions/mean_terminated_length": 0.0,
+      "completions/min_length": 2048.0,
+      "completions/min_terminated_length": 0.0,
+      "epoch": 0.018285714285714287,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23300249874591827,
+      "learning_rate": 7.5e-07,
+      "loss": -0.0,
+      "num_tokens": 2041674.0,
+      "reward": -0.5078557729721069,
+      "reward_std": 0.3458974361419678,
+      "rewards/cosine_scaled_reward/mean": -0.25392788648605347,
+      "rewards/cosine_scaled_reward/std": 0.18378609418869019,
+      "rewards/format_reward/mean": 0.0,
+      "rewards/format_reward/std": 0.0,
+      "step": 16
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1916.0,
+      "completions/mean_length": 1563.734375,
+      "completions/mean_terminated_length": 941.107177734375,
+      "completions/min_length": 389.0,
+      "completions/min_terminated_length": 389.0,
+      "epoch": 0.019428571428571427,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20892462134361267,
+      "learning_rate": 8e-07,
+      "loss": 0.0477,
+      "num_tokens": 2152273.0,
+      "reward": 0.3328002989292145,
+      "reward_std": 0.7669951319694519,
+      "rewards/cosine_scaled_reward/mean": -0.06797486543655396,
+      "rewards/cosine_scaled_reward/std": 0.4412795305252075,
+      "rewards/format_reward/mean": 0.46875,
+      "rewards/format_reward/std": 0.5029674172401428,
+      "step": 17
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1465.0,
+      "completions/mean_length": 1778.90625,
+      "completions/mean_terminated_length": 899.86669921875,
+      "completions/min_length": 535.0,
+      "completions/min_terminated_length": 535.0,
+      "epoch": 0.02057142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.19322611391544342,
+      "learning_rate": 8.499999999999999e-07,
+      "loss": 0.0726,
+      "num_tokens": 2276499.0,
+      "reward": -0.18389344215393066,
+      "reward_std": 0.5934990644454956,
+      "rewards/cosine_scaled_reward/mean": -0.23257172107696533,
+      "rewards/cosine_scaled_reward/std": 0.256833553314209,
+      "rewards/format_reward/mean": 0.28125,
+      "rewards/format_reward/std": 0.4531635046005249,
+      "step": 18
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1771.0,
+      "completions/mean_length": 1869.53125,
+      "completions/mean_terminated_length": 1232.1429443359375,
+      "completions/min_length": 711.0,
+      "completions/min_terminated_length": 711.0,
+      "epoch": 0.021714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21417103707790375,
+      "learning_rate": 9e-07,
+      "loss": 0.0378,
+      "num_tokens": 2407405.0,
+      "reward": -0.05162222683429718,
+      "reward_std": 0.7635236978530884,
+      "rewards/cosine_scaled_reward/mean": -0.158623605966568,
+      "rewards/cosine_scaled_reward/std": 0.4003170132637024,
+      "rewards/format_reward/mean": 0.265625,
+      "rewards/format_reward/std": 0.44515693187713623,
+      "step": 19
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.59375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1836.0,
+      "completions/mean_length": 1572.90625,
+      "completions/mean_terminated_length": 878.5385131835938,
+      "completions/min_length": 369.0,
+      "completions/min_terminated_length": 369.0,
+      "epoch": 0.022857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.1591554582118988,
+      "learning_rate": 9.499999999999999e-07,
+      "loss": 0.0507,
+      "num_tokens": 2519423.0,
+      "reward": 0.2816518545150757,
+      "reward_std": 0.7381908893585205,
+      "rewards/cosine_scaled_reward/mean": -0.07011157274246216,
+      "rewards/cosine_scaled_reward/std": 0.35477158427238464,
+      "rewards/format_reward/mean": 0.421875,
+      "rewards/format_reward/std": 0.49776285886764526,
+      "step": 20
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1827.0,
+      "completions/mean_length": 1776.28125,
+      "completions/mean_terminated_length": 1081.888916015625,
+      "completions/min_length": 332.0,
+      "completions/min_terminated_length": 332.0,
+      "epoch": 0.024,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22487252950668335,
+      "learning_rate": 1e-06,
+      "loss": 0.0137,
+      "num_tokens": 2643913.0,
+      "reward": -0.0122755765914917,
+      "reward_std": 0.4569401443004608,
+      "rewards/cosine_scaled_reward/mean": -0.16238778829574585,
+      "rewards/cosine_scaled_reward/std": 0.3900769054889679,
+      "rewards/format_reward/mean": 0.3125,
+      "rewards/format_reward/std": 0.467176616191864,
+      "step": 21
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.390625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1851.0,
+      "completions/mean_length": 1273.1875,
+      "completions/mean_terminated_length": 776.5128173828125,
+      "completions/min_length": 242.0,
+      "completions/min_terminated_length": 242.0,
+      "epoch": 0.025142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.1901247799396515,
+      "learning_rate": 9.99931462820376e-07,
+      "loss": -0.0442,
+      "num_tokens": 2734413.0,
+      "reward": 0.5235691666603088,
+      "reward_std": 0.4210290312767029,
+      "rewards/cosine_scaled_reward/mean": -0.07415291666984558,
+      "rewards/cosine_scaled_reward/std": 0.40765848755836487,
+      "rewards/format_reward/mean": 0.671875,
+      "rewards/format_reward/std": 0.4732423722743988,
+      "step": 22
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1954.0,
+      "completions/mean_length": 1640.84375,
+      "completions/mean_terminated_length": 1082.888916015625,
+      "completions/min_length": 363.0,
+      "completions/min_terminated_length": 363.0,
+      "epoch": 0.026285714285714287,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21930935978889465,
+      "learning_rate": 9.997258721585931e-07,
+      "loss": 0.0518,
+      "num_tokens": 2850219.0,
+      "reward": 0.23656107485294342,
+      "reward_std": 0.6851356029510498,
+      "rewards/cosine_scaled_reward/mean": -0.10046947002410889,
+      "rewards/cosine_scaled_reward/std": 0.45323267579078674,
+      "rewards/format_reward/mean": 0.4375,
+      "rewards/format_reward/std": 0.5,
+      "step": 23
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1985.0,
+      "completions/mean_length": 1785.265625,
+      "completions/mean_terminated_length": 1113.8333740234375,
+      "completions/min_length": 475.0,
+      "completions/min_terminated_length": 475.0,
+      "epoch": 0.027428571428571427,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.196747824549675,
+      "learning_rate": 9.993832906395582e-07,
+      "loss": 0.0687,
+      "num_tokens": 2975404.0,
+      "reward": 0.04860962927341461,
+      "reward_std": 0.8576602935791016,
+      "rewards/cosine_scaled_reward/mean": -0.1475701779127121,
+      "rewards/cosine_scaled_reward/std": 0.4082482159137726,
+      "rewards/format_reward/mean": 0.34375,
+      "rewards/format_reward/std": 0.4787135720252991,
+      "step": 24
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1752.0,
+      "completions/mean_length": 1695.234375,
+      "completions/mean_terminated_length": 919.1500244140625,
+      "completions/min_length": 502.0,
+      "completions/min_terminated_length": 502.0,
+      "epoch": 0.02857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22251193225383759,
+      "learning_rate": 9.989038226169207e-07,
+      "loss": 0.0401,
+      "num_tokens": 3094195.0,
+      "reward": 0.2244701385498047,
+      "reward_std": 0.6461865901947021,
+      "rewards/cosine_scaled_reward/mean": -0.06745242327451706,
+      "rewards/cosine_scaled_reward/std": 0.41534900665283203,
+      "rewards/format_reward/mean": 0.359375,
+      "rewards/format_reward/std": 0.4836103618144989,
+      "step": 25
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2016.0,
+      "completions/mean_length": 1974.4375,
+      "completions/mean_terminated_length": 1524.888916015625,
+      "completions/min_length": 1105.0,
+      "completions/min_terminated_length": 1105.0,
+      "epoch": 0.029714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23350541293621063,
+      "learning_rate": 9.982876141412855e-07,
+      "loss": 0.0101,
+      "num_tokens": 3231191.0,
+      "reward": 0.16762161254882812,
+      "reward_std": 0.5227605104446411,
+      "rewards/cosine_scaled_reward/mean": -0.041189197450876236,
+      "rewards/cosine_scaled_reward/std": 0.37332749366760254,
+      "rewards/format_reward/mean": 0.25,
+      "rewards/format_reward/std": 0.4364357888698578,
+      "step": 26
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1999.0,
+      "completions/mean_length": 1915.5,
+      "completions/mean_terminated_length": 1277.0909423828125,
+      "completions/min_length": 554.0,
+      "completions/min_terminated_length": 554.0,
+      "epoch": 0.030857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21174418926239014,
+      "learning_rate": 9.975348529157229e-07,
+      "loss": 0.0312,
+      "num_tokens": 3364071.0,
+      "reward": -0.18293717503547668,
+      "reward_std": 0.5386844873428345,
+      "rewards/cosine_scaled_reward/mean": -0.20865610241889954,
+      "rewards/cosine_scaled_reward/std": 0.2562413811683655,
+      "rewards/format_reward/mean": 0.234375,
+      "rewards/format_reward/std": 0.42695629596710205,
+      "step": 27
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2007.0,
+      "completions/mean_length": 1815.140625,
+      "completions/mean_terminated_length": 1220.0555419921875,
+      "completions/min_length": 445.0,
+      "completions/min_terminated_length": 445.0,
+      "epoch": 0.032,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.213092640042305,
+      "learning_rate": 9.96645768238595e-07,
+      "loss": 0.0361,
+      "num_tokens": 3490576.0,
+      "reward": 0.04266031086444855,
+      "reward_std": 0.776748776435852,
+      "rewards/cosine_scaled_reward/mean": -0.13491985201835632,
+      "rewards/cosine_scaled_reward/std": 0.37269750237464905,
+      "rewards/format_reward/mean": 0.3125,
+      "rewards/format_reward/std": 0.467176616191864,
+      "step": 28
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2033.0,
+      "completions/mean_length": 1906.15625,
+      "completions/mean_terminated_length": 1039.3333740234375,
+      "completions/min_length": 633.0,
+      "completions/min_terminated_length": 633.0,
+      "epoch": 0.03314285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22322852909564972,
+      "learning_rate": 9.956206309337066e-07,
+      "loss": 0.0389,
+      "num_tokens": 3623042.0,
+      "reward": -0.1004815474152565,
+      "reward_std": 0.539789080619812,
+      "rewards/cosine_scaled_reward/mean": -0.12836576998233795,
+      "rewards/cosine_scaled_reward/std": 0.28681084513664246,
+      "rewards/format_reward/mean": 0.15625,
+      "rewards/format_reward/std": 0.36596253514289856,
+      "step": 29
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1992.0,
+      "completions/mean_length": 1854.1875,
+      "completions/mean_terminated_length": 1162.0,
+      "completions/min_length": 592.0,
+      "completions/min_terminated_length": 592.0,
+      "epoch": 0.03428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2063974291086197,
+      "learning_rate": 9.944597532678119e-07,
+      "loss": 0.0115,
+      "num_tokens": 3752246.0,
+      "reward": -0.030107807368040085,
+      "reward_std": 0.6322507858276367,
+      "rewards/cosine_scaled_reward/mean": -0.1634913980960846,
+      "rewards/cosine_scaled_reward/std": 0.31110286712646484,
+      "rewards/format_reward/mean": 0.296875,
+      "rewards/format_reward/std": 0.4604927599430084,
+      "step": 30
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.84375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1523.0,
+      "completions/mean_length": 1841.15625,
+      "completions/mean_terminated_length": 724.2000122070312,
+      "completions/min_length": 165.0,
+      "completions/min_terminated_length": 165.0,
+      "epoch": 0.03542857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2025275081396103,
+      "learning_rate": 9.931634888554935e-07,
+      "loss": 0.0143,
+      "num_tokens": 3880576.0,
+      "reward": -0.34719598293304443,
+      "reward_std": 0.5259275436401367,
+      "rewards/cosine_scaled_reward/mean": -0.2595354914665222,
+      "rewards/cosine_scaled_reward/std": 0.24079306423664093,
+      "rewards/format_reward/mean": 0.171875,
+      "rewards/format_reward/std": 0.38025420904159546,
+      "step": 31
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.84375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1998.0,
+      "completions/mean_length": 1945.65625,
+      "completions/mean_terminated_length": 1393.0,
+      "completions/min_length": 899.0,
+      "completions/min_terminated_length": 899.0,
+      "epoch": 0.036571428571428574,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22421319782733917,
+      "learning_rate": 9.917322325514487e-07,
+      "loss": 0.0542,
+      "num_tokens": 4015450.0,
+      "reward": -0.2238868921995163,
+      "reward_std": 0.6127103567123413,
+      "rewards/cosine_scaled_reward/mean": -0.20569345355033875,
+      "rewards/cosine_scaled_reward/std": 0.26141345500946045,
+      "rewards/format_reward/mean": 0.1875,
+      "rewards/format_reward/std": 0.39339789748191833,
+      "step": 32
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.90625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1959.0,
+      "completions/mean_length": 1976.890625,
+      "completions/mean_terminated_length": 1289.5,
+      "completions/min_length": 581.0,
+      "completions/min_terminated_length": 581.0,
+      "epoch": 0.037714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2219865769147873,
+      "learning_rate": 9.901664203302124e-07,
+      "loss": 0.0139,
+      "num_tokens": 4153187.0,
+      "reward": -0.5050230026245117,
+      "reward_std": 0.38754361867904663,
+      "rewards/cosine_scaled_reward/mean": -0.31501150131225586,
+      "rewards/cosine_scaled_reward/std": 0.19765734672546387,
+      "rewards/format_reward/mean": 0.125,
+      "rewards/format_reward/std": 0.3333333432674408,
+      "step": 33
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.515625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2029.0,
+      "completions/mean_length": 1573.625,
+      "completions/mean_terminated_length": 1068.6451416015625,
+      "completions/min_length": 517.0,
+      "completions/min_terminated_length": 517.0,
+      "epoch": 0.038857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23140408098697662,
+      "learning_rate": 9.88466529153356e-07,
+      "loss": 0.0697,
+      "num_tokens": 4263451.0,
+      "reward": 0.3802332282066345,
+      "reward_std": 0.8625352382659912,
+      "rewards/cosine_scaled_reward/mean": -0.05207090824842453,
+      "rewards/cosine_scaled_reward/std": 0.4423771798610687,
+      "rewards/format_reward/mean": 0.484375,
+      "rewards/format_reward/std": 0.5037065148353577,
+      "step": 34
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1957.0,
+      "completions/mean_length": 1814.875,
+      "completions/mean_terminated_length": 1053.3333740234375,
+      "completions/min_length": 359.0,
+      "completions/min_terminated_length": 359.0,
+      "epoch": 0.04,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21747314929962158,
+      "learning_rate": 9.866330768241983e-07,
+      "loss": 0.0288,
+      "num_tokens": 4391099.0,
+      "reward": 0.11022068560123444,
+      "reward_std": 0.898347795009613,
+      "rewards/cosine_scaled_reward/mean": -0.08551465719938278,
+      "rewards/cosine_scaled_reward/std": 0.4119128882884979,
+      "rewards/format_reward/mean": 0.28125,
+      "rewards/format_reward/std": 0.4531635046005249,
+      "step": 35
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.90625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1775.0,
+      "completions/mean_length": 1976.765625,
+      "completions/mean_terminated_length": 1288.166748046875,
+      "completions/min_length": 964.0,
+      "completions/min_terminated_length": 964.0,
+      "epoch": 0.04114285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23834578692913055,
+      "learning_rate": 9.846666218300807e-07,
+      "loss": 0.021,
+      "num_tokens": 4528724.0,
+      "reward": -0.38736510276794434,
+      "reward_std": 0.5356569290161133,
+      "rewards/cosine_scaled_reward/mean": -0.24837006628513336,
+      "rewards/cosine_scaled_reward/std": 0.23275430500507355,
+      "rewards/format_reward/mean": 0.109375,
+      "rewards/format_reward/std": 0.3145764470100403,
+      "step": 36
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.9375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2047.0,
+      "completions/mean_length": 2013.9375,
+      "completions/mean_terminated_length": 1503.0,
+      "completions/min_length": 1027.0,
+      "completions/min_terminated_length": 1027.0,
+      "epoch": 0.04228571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22654284536838531,
+      "learning_rate": 9.825677631722435e-07,
+      "loss": 0.0142,
+      "num_tokens": 4668640.0,
+      "reward": -0.42377781867980957,
+      "reward_std": 0.379480242729187,
+      "rewards/cosine_scaled_reward/mean": -0.2665764391422272,
+      "rewards/cosine_scaled_reward/std": 0.18001720309257507,
+      "rewards/format_reward/mean": 0.109375,
+      "rewards/format_reward/std": 0.3145764470100403,
+      "step": 37
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1254.0,
+      "completions/mean_length": 1912.265625,
+      "completions/mean_terminated_length": 1082.77783203125,
+      "completions/min_length": 920.0,
+      "completions/min_terminated_length": 920.0,
+      "epoch": 0.04342857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22495149075984955,
+      "learning_rate": 9.80337140183366e-07,
+      "loss": 0.0353,
+      "num_tokens": 4802737.0,
+      "reward": -0.15185467898845673,
+      "reward_std": 0.38927191495895386,
+      "rewards/cosine_scaled_reward/mean": -0.14623984694480896,
+      "rewards/cosine_scaled_reward/std": 0.32866883277893066,
+      "rewards/format_reward/mean": 0.140625,
+      "rewards/format_reward/std": 0.3503824472427368,
+      "step": 38
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1979.0,
+      "completions/mean_length": 1678.4375,
+      "completions/mean_terminated_length": 656.7058715820312,
+      "completions/min_length": 300.0,
+      "completions/min_terminated_length": 300.0,
+      "epoch": 0.044571428571428574,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.19435559213161469,
+      "learning_rate": 9.779754323328192e-07,
+      "loss": 0.0416,
+      "num_tokens": 4920941.0,
+      "reward": 0.17510981857776642,
+      "reward_std": 0.559760570526123,
+      "rewards/cosine_scaled_reward/mean": -0.0765075832605362,
+      "rewards/cosine_scaled_reward/std": 0.3369429409503937,
+      "rewards/format_reward/mean": 0.328125,
+      "rewards/format_reward/std": 0.4732423722743988,
+      "step": 39
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1762.0,
+      "completions/mean_length": 1682.25,
+      "completions/mean_terminated_length": 877.6000366210938,
+      "completions/min_length": 465.0,
+      "completions/min_terminated_length": 465.0,
+      "epoch": 0.045714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.19440439343452454,
+      "learning_rate": 9.754833590196926e-07,
+      "loss": 0.0685,
+      "num_tokens": 5038677.0,
+      "reward": 0.09382888674736023,
+      "reward_std": 0.4140171706676483,
+      "rewards/cosine_scaled_reward/mean": -0.12496057152748108,
+      "rewards/cosine_scaled_reward/std": 0.3649806082248688,
+      "rewards/format_reward/mean": 0.34375,
+      "rewards/format_reward/std": 0.4787135720252991,
+      "step": 40
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1905.0,
+      "completions/mean_length": 1841.0625,
+      "completions/mean_terminated_length": 1417.3333740234375,
+      "completions/min_length": 965.0,
+      "completions/min_terminated_length": 965.0,
+      "epoch": 0.046857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21685408055782318,
+      "learning_rate": 9.728616793536587e-07,
+      "loss": 0.0105,
+      "num_tokens": 5167657.0,
+      "reward": -0.15476089715957642,
+      "reward_std": 0.5854519605636597,
+      "rewards/cosine_scaled_reward/mean": -0.2648804187774658,
+      "rewards/cosine_scaled_reward/std": 0.26939424872398376,
+      "rewards/format_reward/mean": 0.375,
+      "rewards/format_reward/std": 0.48795005679130554,
+      "step": 41
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1933.0,
+      "completions/mean_length": 1695.0,
+      "completions/mean_terminated_length": 719.058837890625,
+      "completions/min_length": 205.0,
+      "completions/min_terminated_length": 205.0,
+      "epoch": 0.048,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23794473707675934,
+      "learning_rate": 9.701111919237408e-07,
+      "loss": 0.0327,
+      "num_tokens": 5286497.0,
+      "reward": -0.2923233211040497,
+      "reward_std": 0.36149862408638,
+      "rewards/cosine_scaled_reward/mean": -0.27897417545318604,
+      "rewards/cosine_scaled_reward/std": 0.17192503809928894,
+      "rewards/format_reward/mean": 0.265625,
+      "rewards/format_reward/std": 0.44515693187713623,
+      "step": 42
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2045.0,
+      "completions/mean_length": 1793.84375,
+      "completions/mean_terminated_length": 1031.375,
+      "completions/min_length": 714.0,
+      "completions/min_terminated_length": 714.0,
+      "epoch": 0.04914285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21354877948760986,
+      "learning_rate": 9.672327345550543e-07,
+      "loss": 0.0597,
+      "num_tokens": 5412919.0,
+      "reward": -0.0004070103168487549,
+      "reward_std": 0.5297929048538208,
+      "rewards/cosine_scaled_reward/mean": -0.12520350515842438,
+      "rewards/cosine_scaled_reward/std": 0.3128352463245392,
+      "rewards/format_reward/mean": 0.25,
+      "rewards/format_reward/std": 0.4364357888698578,
+      "step": 43
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1729.0,
+      "completions/mean_length": 1651.1875,
+      "completions/mean_terminated_length": 778.2000122070312,
+      "completions/min_length": 251.0,
+      "completions/min_terminated_length": 251.0,
+      "epoch": 0.05028571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20100052654743195,
+      "learning_rate": 9.64227184053598e-07,
+      "loss": 0.0089,
+      "num_tokens": 5529291.0,
+      "reward": 0.13101597130298615,
+      "reward_std": 0.5976744890213013,
+      "rewards/cosine_scaled_reward/mean": -0.09855452179908752,
+      "rewards/cosine_scaled_reward/std": 0.46046286821365356,
+      "rewards/format_reward/mean": 0.328125,
+      "rewards/format_reward/std": 0.4732423722743988,
+      "step": 44
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1953.0,
+      "completions/mean_length": 2011.765625,
+      "completions/mean_terminated_length": 1584.2000732421875,
+      "completions/min_length": 1146.0,
+      "completions/min_terminated_length": 1146.0,
+      "epoch": 0.05142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2144525796175003,
+      "learning_rate": 9.610954559391704e-07,
+      "loss": 0.0108,
+      "num_tokens": 5669700.0,
+      "reward": -0.15992262959480286,
+      "reward_std": 0.5183610916137695,
+      "rewards/cosine_scaled_reward/mean": -0.14246131479740143,
+      "rewards/cosine_scaled_reward/std": 0.37169432640075684,
+      "rewards/format_reward/mean": 0.125,
+      "rewards/format_reward/std": 0.3333333432674408,
+      "step": 45
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1566.0,
+      "completions/mean_length": 1811.21875,
+      "completions/mean_terminated_length": 965.5714721679688,
+      "completions/min_length": 578.0,
+      "completions/min_terminated_length": 578.0,
+      "epoch": 0.052571428571428575,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2243409901857376,
+      "learning_rate": 9.578385041664925e-07,
+      "loss": 0.0324,
+      "num_tokens": 5796786.0,
+      "reward": -0.2682954668998718,
+      "reward_std": 0.47855472564697266,
+      "rewards/cosine_scaled_reward/mean": -0.2435227334499359,
+      "rewards/cosine_scaled_reward/std": 0.21708372235298157,
+      "rewards/format_reward/mean": 0.21875,
+      "rewards/format_reward/std": 0.4166666865348816,
+      "step": 46
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1730.0,
+      "completions/mean_length": 1720.15625,
+      "completions/mean_terminated_length": 882.3333129882812,
+      "completions/min_length": 432.0,
+      "completions/min_terminated_length": 432.0,
+      "epoch": 0.053714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.1925242692232132,
+      "learning_rate": 9.54457320834625e-07,
+      "loss": 0.0641,
+      "num_tokens": 5917276.0,
+      "reward": -0.03124237060546875,
+      "reward_std": 0.6693180203437805,
+      "rewards/cosine_scaled_reward/mean": -0.17968368530273438,
+      "rewards/cosine_scaled_reward/std": 0.379862517118454,
+      "rewards/format_reward/mean": 0.328125,
+      "rewards/format_reward/std": 0.4732423722743988,
+      "step": 47
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2012.0,
+      "completions/mean_length": 1723.734375,
+      "completions/mean_terminated_length": 1059.761962890625,
+      "completions/min_length": 617.0,
+      "completions/min_terminated_length": 617.0,
+      "epoch": 0.054857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20502391457557678,
+      "learning_rate": 9.509529358847654e-07,
+      "loss": 0.0544,
+      "num_tokens": 6038139.0,
+      "reward": 0.21815460920333862,
+      "reward_std": 0.6701791286468506,
+      "rewards/cosine_scaled_reward/mean": -0.05498518794775009,
+      "rewards/cosine_scaled_reward/std": 0.42852458357810974,
+      "rewards/format_reward/mean": 0.328125,
+      "rewards/format_reward/std": 0.4732423722743988,
+      "step": 48
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.515625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1571.0,
+      "completions/mean_length": 1450.3125,
+      "completions/mean_terminated_length": 814.0645141601562,
+      "completions/min_length": 275.0,
+      "completions/min_terminated_length": 275.0,
+      "epoch": 0.056,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.18392837047576904,
+      "learning_rate": 9.473264167865171e-07,
+      "loss": 0.0771,
+      "num_tokens": 6141023.0,
+      "reward": 0.20156216621398926,
+      "reward_std": 0.7049944400787354,
+      "rewards/cosine_scaled_reward/mean": -0.14921891689300537,
+      "rewards/cosine_scaled_reward/std": 0.35212206840515137,
+      "rewards/format_reward/mean": 0.5,
+      "rewards/format_reward/std": 0.5039526224136353,
+      "step": 49
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1605.0,
+      "completions/mean_length": 1740.84375,
+      "completions/mean_terminated_length": 643.857177734375,
+      "completions/min_length": 307.0,
+      "completions/min_terminated_length": 307.0,
+      "epoch": 0.05714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.1936425119638443,
+      "learning_rate": 9.43578868212728e-07,
+      "loss": 0.0292,
+      "num_tokens": 6263253.0,
+      "reward": -0.08827750384807587,
+      "reward_std": 0.3788633346557617,
+      "rewards/cosine_scaled_reward/mean": -0.16132624447345734,
+      "rewards/cosine_scaled_reward/std": 0.36572694778442383,
+      "rewards/format_reward/mean": 0.234375,
+      "rewards/format_reward/std": 0.42695629596710205,
+      "step": 50
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 200,
+  "num_input_tokens_seen": 6263253,
+  "num_train_epochs": 1,
+  "save_steps": 50,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-50/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bfcdb0ec897893bf3e1232fa71465d92c0536fcf2ace560791d47a887fbdd3c
+size 8888

checkpoint-50/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)