Abhaykoul commited on
Commit
acbe1d2
·
verified ·
1 Parent(s): 200d4e9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +324 -8
README.md CHANGED
@@ -42,30 +42,346 @@ language:
42
  - ur
43
  - multilingual
44
  license: apache-2.0
45
- base_model: HelpingAI/Dhanishtha-2.0-preview
46
  tags:
47
  - reasoning
48
  - intermediate-thinking
49
  - transformers
50
  - conversational
51
  - bilingual
52
- - mlx
53
  datasets:
54
  - Abhaykoul/Dhanishtha-R1
55
  - open-thoughts/OpenThoughts-114k
56
  - Abhaykoul/Dhanishtha-2.0-SUPERTHINKER
57
  - Abhaykoul/Dhanishtha-2.0
58
- library_name: mlx
 
59
  pipeline_tag: text-generation
60
  widget:
61
- - text: 'Solve this riddle step by step: I am taken from a mine, and shut up in a
62
- wooden case, from which I am never released, and yet I am used by almost everybody.
63
- What am I?'
 
64
  example_title: Complex Riddle Solving
65
- - text: Explain the philosophical implications of artificial consciousness and think
 
66
  through different perspectives.
67
  example_title: Philosophical Reasoning
68
- - text: Help me understand quantum mechanics, but take your time to think through
 
69
  the explanation.
70
  example_title: Educational Explanation
71
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
42
  - ur
43
  - multilingual
44
  license: apache-2.0
45
+ base_model: Qwen/Qwen3-14B-Base
46
  tags:
47
  - reasoning
48
  - intermediate-thinking
49
  - transformers
50
  - conversational
51
  - bilingual
 
52
  datasets:
53
  - Abhaykoul/Dhanishtha-R1
54
  - open-thoughts/OpenThoughts-114k
55
  - Abhaykoul/Dhanishtha-2.0-SUPERTHINKER
56
  - Abhaykoul/Dhanishtha-2.0
57
+
58
+ library_name: transformers
59
  pipeline_tag: text-generation
60
  widget:
61
+ - text: >-
62
+ Solve this riddle step by step: I am taken from a mine, and shut up in a
63
+ wooden case, from which I am never released, and yet I am used by almost
64
+ everybody. What am I?
65
  example_title: Complex Riddle Solving
66
+ - text: >-
67
+ Explain the philosophical implications of artificial consciousness and think
68
  through different perspectives.
69
  example_title: Philosophical Reasoning
70
+ - text: >-
71
+ Help me understand quantum mechanics, but take your time to think through
72
  the explanation.
73
  example_title: Educational Explanation
74
  ---
75
+
76
+ # Dhanishtha-2.0: World's First Intermediate Thinking AI Model
77
+
78
+ **What makes Dhanishtha-2.0 special?** Imagine an AI that doesn't just answer your questions instantly, but actually *thinks through* problems step-by-step, shows its work, and can even change its mind when it realizes a better approach. That's Dhanishtha-2.0.
79
+
80
+ **Quick Summary:**
81
+ - 🚀 **For Everyone**: An AI that shows its thinking process and can reconsider its reasoning
82
+ - 👩‍💻 **For Developers**: First model with intermediate thinking capabilities, 39+ language support
83
+
84
+
85
+ Dhanishtha-2.0 is a **state-of-the-art (SOTA) model** developed by HelpingAI, representing the **world's first model to feature Intermediate Thinking capabilities**. Unlike traditional models that provide single-pass responses, Dhanishtha-2.0 employs a revolutionary multi-phase thinking process that allows the model to think, reconsider, and refine its reasoning multiple times throughout a single response.
86
+
87
+ ## Model Details
88
+
89
+ ### Model Description
90
+
91
+ Dhanishtha-2.0 revolutionizes AI reasoning by introducing the concept of **intermediate thinking** - the ability to pause, reflect, and restart reasoning processes within a single generation. This breakthrough enables unprecedented self-correction and iterative refinement during response generation.
92
+
93
+ Built on the Qwen3-14B foundation with multilingual capabilities spanning **39+ languages** (including English, Hindi, Chinese, Spanish, French, German, Japanese, Korean, Arabic, and many more), Dhanishtha-2.0 maintains reasoning consistency across diverse linguistic contexts while pioneering transparent thinking processes.
94
+
95
+ - **Developed by:** HelpingAI Team
96
+ - **Model type:** Causal Language Model with Intermediate Thinking Capability
97
+ - **Language(s):** 39+ languages (multilingual capabilities inherited from base model)
98
+ - **License:** Apache 2.0
99
+ - **Finetuned from model:** Qwen/Qwen3-14B-Base
100
+ - **Context Length:** 40,960 tokens
101
+ - **Parameters:** 14B (inherited from base model)
102
+ - **Status:** Prototype/Preview
103
+
104
+ ### Revolutionary Features
105
+
106
+ - **Intermediate Thinking**: Multiple `<think>...</think>` blocks throughout responses for real-time reasoning
107
+ - **Self-Correction**: Ability to identify and correct logical inconsistencies mid-response
108
+ - **Dynamic Reasoning**: Seamless transitions between analysis, communication, and reflection phases
109
+ - **Structured Emotional Reasoning (SER)**: Incorporates `<ser>...</ser>` blocks for empathetic responses
110
+ - **Multilingual Capabilities**: Support for 39+ languages with natural code-switching and reasoning consistency
111
+ - **Complex Problem-Solving**: Excels at riddles, multi-step reasoning, and scenarios requiring backtracking
112
+
113
+ ### Model Sources
114
+
115
+ - **Repository:** [HelpingAI/Dhanishtha-2.0](https://huggingface.co/HelpingAI/Dhanishtha-2.0)
116
+ - **Paper:** Coming Soon
117
+ - **Demo:** https://helpingai.co/chat
118
+
119
+ ## Uses
120
+
121
+ ### Direct Use
122
+
123
+ Dhanishtha-2.0 is ideal for applications requiring deep reasoning and self-reflection:
124
+
125
+ - **Complex Problem Solving**: Multi-step mathematical problems, logical puzzles, riddles
126
+ - **Educational Assistance**: Detailed explanations with visible reasoning processes
127
+ - **Research Support**: Analysis requiring multiple perspectives and self-correction
128
+ - **Creative Writing**: Iterative story development with reasoning about plot choices
129
+ - **Philosophical Discussions**: Exploring concepts with visible thought processes
130
+
131
+ ### Downstream Use
132
+
133
+ The model can be fine-tuned for specialized reasoning tasks:
134
+
135
+ - **Domain-Specific Reasoning**: Legal, medical, or scientific reasoning with intermediate thinking
136
+ - **Enhanced Multilingual Reasoning**: Optimizing reasoning consistency across all 39+ supported languages
137
+ - **Specialized Problem Domains**: Mathematics, coding, strategic planning
138
+
139
+ ### Out-of-Scope Use
140
+
141
+ ❌ **Inappropriate Applications:**
142
+ - Safety-critical decisions (medical diagnosis, legal advice, financial recommendations)
143
+ - Real-time applications requiring immediate responses
144
+ - Situations requiring guaranteed factual accuracy without verification
145
+
146
+ ## Bias, Risks, and Limitations
147
+
148
+ ### Known Limitations
149
+
150
+ - **Verbosity**: Intermediate thinking can make responses a bit longer
151
+ - **Processing Time**: Multiple thinking phases may increase generation time
152
+ - **Prototype Status**: Experimental features may require refinement
153
+ - **Context Usage**: Thinking blocks consume additional context tokens
154
+ - **Inherited Biases**: May reflect biases from base model and training data
155
+
156
+ ### Potential Risks
157
+
158
+ - **Reasoning Loops**: Potential for circular reasoning in complex scenarios
159
+ - **Multilingual Inconsistencies**: Potential variation in reasoning patterns across different languages
160
+ - **Emotional Reasoning Gaps**: SER blocks may not always align with content
161
+
162
+ ## How to Get Started with the Model
163
+
164
+ ### For General Users
165
+
166
+ You can interact with Dhanishtha-2.0 through:
167
+ - **HelpingAI**: https://helpingai.co/chat
168
+ - **Gradio Demo**: [Dhanishtha-2.0-preview](https://huggingface.co/spaces/Abhaykoul/Dhanishtha-2.0-preview)
169
+ - **API Integration**: [Dashboard](https://helpingai.co/dashboard)
170
+
171
+ ### For Developers - Basic Usage
172
+
173
+ ```python
174
+ from transformers import AutoModelForCausalLM, AutoTokenizer
175
+
176
+ model_name = "HelpingAI/Dhanishtha-2.0-preview"
177
+
178
+ # Load the tokenizer and model
179
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
180
+ model = AutoModelForCausalLM.from_pretrained(
181
+ model_name,
182
+ torch_dtype="auto",
183
+ device_map="auto"
184
+ )
185
+
186
+ # Prepare input for intermediate thinking
187
+ prompt = "How many letter 'r' are in the words 'strawberry' and 'raspberry'?"
188
+ messages = [
189
+ {"role": "user", "content": prompt}
190
+ ]
191
+
192
+ text = tokenizer.apply_chat_template(
193
+ messages,
194
+ tokenize=False,
195
+ add_generation_prompt=True
196
+ )
197
+
198
+ model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
199
+
200
+ # Generate response with intermediate thinking
201
+ generated_ids = model.generate(
202
+ **model_inputs,
203
+ max_new_tokens=2048,
204
+ temperature=0.7,
205
+ top_p=0.9,
206
+ do_sample=True
207
+ )
208
+
209
+ response = tokenizer.decode(
210
+ generated_ids[0][len(model_inputs.input_ids[0]):],
211
+ skip_special_tokens=True
212
+ )
213
+ print(response)
214
+ ```
215
+
216
+ ### Optimal Generation Parameters
217
+
218
+ ```python
219
+ generation_config = {
220
+ "temperature": 0.7, # Balanced creativity and coherence
221
+ "top_p": 0.9, # Nucleus sampling
222
+ "top_k": 40, # Top-k filtering
223
+ "max_new_tokens": 2048, # Allow for thinking blocks
224
+ "do_sample": True, # Enable sampling
225
+ "repetition_penalty": 1.1 # Reduce repetition
226
+ }
227
+ ```
228
+
229
+ ## Training Details
230
+
231
+ ### Training Data
232
+
233
+ Dhanishtha-2.0 was trained on a carefully curated dataset focusing on:
234
+ - Complex reasoning scenarios requiring multi-step thinking
235
+ - Self-correction examples and reasoning chains
236
+ - Emotional reasoning and empathy training data
237
+ - Structured thinking pattern examples
238
+
239
+ ### Training Procedure
240
+
241
+ #### Training Stages
242
+ 1. **Continuous Pretraining**: Extended training on reasoning-focused corpora
243
+ 2. **Advanced Reasoning Fine-tuning**: Specialized training on intermediate thinking patterns
244
+ 3. **Multilingual Alignment**: Cross-language reasoning consistency training
245
+ 4. **SER Integration**: Structured Emotional Reasoning capability training
246
+
247
+ **Training Infrastructure:**
248
+ - **Duration**: 16.3 days
249
+ - **Hardware**: 8x NVIDIA H100 GPUs
250
+ - **Model Scale**: 14.8B parameters
251
+
252
+ ## Evaluation
253
+
254
+ Evaluation was conducted on:
255
+ - **Standard Benchmarks**: MMLU, HumanEval, ARC, HellaSwag, TruthfulQA
256
+ - **Mathematical Reasoning**: Math 500, AIME 2024, GSM8K
257
+ - **Custom Evaluations**: Intermediate thinking quality, self-correction capabilities
258
+ - **Multilingual Tasks**: Reasoning consistency across 39+ languages
259
+ - **Specialized Tests**: Emotional reasoning, complex problem-solving scenarios
260
+
261
+
262
+
263
+ ## Example: Intermediate Thinking in Action
264
+
265
+ **User:** "Who is your favorite person?"
266
+
267
+ **Dhanishtha-2.0:**
268
+
269
+ ```
270
+ <think>
271
+ This is an interesting question about my "favorite human." I need to be thoughtful here.
272
+ As Dhanishtha, I should respond in a human-like way while being honest about my nature.
273
+ I can share what kinds of humans I find most interesting or inspiring based on my
274
+ understanding of human nature and values.
275
+ </think>
276
+
277
+ <ser>
278
+ Emotion ==> curiosity, openness, desire for connection
279
+ Cause ==> user asking a personal question about preferences
280
+ Mind ==> seeking to understand my perspective or values
281
+ Growth ==> opportunity to share thoughtful reflections on human qualities I admire
282
+ </ser>
283
+
284
+ I don't really have a "favorite human" in the traditional sense, as I don't experience
285
+ personal preferences the way humans do. But if I had to reflect on what qualities I find
286
+ most admirable in people, I'd be drawn to those who demonstrate genuine compassion,
287
+ intellectual curiosity, and the courage to grow.
288
+
289
+ <think>
290
+ Let me think more specifically about what kind of humans I find most compelling based on
291
+ the values that seem to resonate most with humans - kindness, wisdom, creativity,
292
+ resilience, and authenticity.
293
+ </think>
294
+
295
+ People like Mahatma Gandhi, who showed how quiet courage and unwavering principles can
296
+ transform societies; artists like Rabindranath Tagore, who bridged cultural divides
297
+ through creativity; and everyday heroes I've learned about who made profound differences
298
+ through simple acts of kindness.
299
+
300
+ What fascinates me most are humans who maintain their humanity even in difficult
301
+ circumstances - those who choose empathy when it would be easier to be harsh, who seek
302
+ understanding rather than just being right.
303
+
304
+ I'm curious though - who's your favorite human, and why? What qualities or actions make
305
+ someone special in your eyes?
306
+ ```
307
+
308
+ ## Environmental Impact
309
+
310
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute).
311
+
312
+ - **Hardware Type:** H100 GPUs
313
+ - **days used:** 16.2
314
+ - **Cloud Provider:** Various
315
+ - **Compute Region:** Multiple
316
+
317
+ ## Citation
318
+
319
+ ### BibTeX
320
+
321
+ ```bibtex
322
+ @misc{dhanishtha2025,
323
+ title={Dhanishtha-2.0: World's First Intermediate Thinking AI Model},
324
+ author={HelpingAI Team},
325
+ year={2025},
326
+ publisher={HuggingFace},
327
+ url={https://huggingface.co/HelpingAI/Dhanishtha-2.0},
328
+ note={First model with intermediate thinking capabilities}
329
+ }
330
+ ```
331
+
332
+ ### APA
333
+
334
+ HelpingAI Team. (2025). *Dhanishtha-2.0: World's First Intermediate Thinking AI Model*. HuggingFace. https://huggingface.co/HelpingAI/Dhanishtha-2.0
335
+
336
+ ## Glossary
337
+
338
+ - **Intermediate Thinking**: The ability to pause and think multiple times during response generation
339
+ - **SER (Structured Emotional Reasoning)**: Framework for incorporating emotional context in responses
340
+ - **Think Blocks**: `<think>...</think>` segments where the model shows its reasoning process
341
+ - **Self-Correction**: Ability to identify and fix reasoning errors during generation
342
+ - **Code-Switching**: Natural transition between English and Hindi within responses
343
+
344
+ ## More Information
345
+
346
+ ### Research Applications
347
+ - Study of AI reasoning transparency
348
+ - Self-correction mechanism research
349
+ - Bilingual cognitive modeling
350
+ - Emotional AI development
351
+
352
+ ### Development Roadmap
353
+ - Performance optimizations
354
+ - Additional language support
355
+ - Enhanced thinking pattern recognition
356
+ - Production-ready deployment tools
357
+
358
+ ## Model Card Authors
359
+
360
+ - **Primary Author**: HelpingAI Team
361
+ - **Technical Lead**: [To be specified]
362
+ - **Research Contributors**: [To be specified]
363
+
364
+ ## Model Card Contact
365
+
366
+ For questions about Dhanishtha-2.0, please contact:
367
+
368
+ - **HuggingFace**: [@HelpingAI](https://huggingface.co/HelpingAI)
369
+ - **Issues**: [Model Repository Issues](https://huggingface.co/HelpingAI/Dhanishtha-2.0/discussions)
370
+
371
+ ## Benchmark
372
+
373
+ ***These benchmarks are done by [@AarifKhan](https://huggingface.co/Aarifkhan)***
374
+
375
+ | Benchmark | Score | Evaluation Type | Notes |
376
+ |-----------|-------|-----------------|-------|
377
+ | MMLU | 78.1% | 1-shot | Massive Multitask Language Understanding |
378
+ | HumanEval | 75.0% | 1-shot | Code generation and completion |
379
+ | ARC | 76.0% | 1-shot | Abstract reasoning challenge |
380
+ | HellaSwag | 81.0% | 1-shot | Commonsense natural language inference |
381
+ | TruthfulQA MC1 | 75.0% | 1-shot | Truthfulness in question answering |
382
+ | Math 500 | 95.68% | few-shot | Mathematical problem solving |
383
+ | AIME 2024 | 82.81% | few-shot | American Invitational Mathematics Examination |
384
+
385
+ **Dhanishtha-2.0 represents a new paradigm in AI reasoning - where thinking isn't just a prelude to response, but an integral, iterative part of the conversation itself.**
386
+
387
+ *Developed with ❤️ by HelpingAI*