Datasets:
Tasks:
Text Retrieval
Modalities:
Text
Formats:
parquet
Sub-tasks:
document-retrieval
Languages:
Japanese
Size:
< 1K
License:
Upload folder using huggingface_hub
Browse files- README.md +67 -0
- corpus/corpus-00000-of-00001.parquet +3 -0
- create_sample_dataset.py +116 -0
- data/test-00000-of-00001.parquet +3 -0
- queries/queries-00000-of-00001.parquet +3 -0
README.md
ADDED
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
annotations_creators:
|
3 |
+
- derived
|
4 |
+
language:
|
5 |
+
- ja
|
6 |
+
license: cc-by-sa-4.0
|
7 |
+
multilinguality: monolingual
|
8 |
+
task_categories:
|
9 |
+
- text-retrieval
|
10 |
+
task_ids:
|
11 |
+
- document-retrieval
|
12 |
+
tags:
|
13 |
+
- mteb
|
14 |
+
- text
|
15 |
+
- code
|
16 |
+
- japanese
|
17 |
+
- sample
|
18 |
+
---
|
19 |
+
|
20 |
+
# JapaneseCode1Retrieval-sample
|
21 |
+
|
22 |
+
A sample dataset for Japanese-English code retrieval evaluation. This dataset contains Japanese natural language descriptions paired with Python code snippets.
|
23 |
+
|
24 |
+
## Task category
|
25 |
+
|
26 |
+
Retrieval
|
27 |
+
|
28 |
+
## Domains
|
29 |
+
|
30 |
+
Programming, Code Generation
|
31 |
+
|
32 |
+
## Dataset Structure
|
33 |
+
|
34 |
+
The dataset follows the standard MTEB retrieval format:
|
35 |
+
|
36 |
+
- `corpus/corpus-00000-of-00001.parquet`: 5 Python code documents with fields `_id`, `title`, `text`
|
37 |
+
- `queries/queries-00000-of-00001.parquet`: 5 Japanese queries with fields `_id`, `text`
|
38 |
+
- `data/test-00000-of-00001.parquet`: 5 relevance judgments with fields `query-id`, `corpus-id`, `score`
|
39 |
+
|
40 |
+
## Usage
|
41 |
+
|
42 |
+
You can evaluate an embedding model on this sample dataset using the following code:
|
43 |
+
|
44 |
+
```python
|
45 |
+
import mteb
|
46 |
+
|
47 |
+
# Load the sample dataset
|
48 |
+
task = mteb.get_task("JapaneseCode1Retrieval-sample")
|
49 |
+
evaluator = mteb.MTEB(tasks=[task])
|
50 |
+
|
51 |
+
# Run evaluation with your model
|
52 |
+
model = mteb.get_model("your-model-name")
|
53 |
+
results = evaluator.run(model)
|
54 |
+
```
|
55 |
+
|
56 |
+
## Sample Content
|
57 |
+
|
58 |
+
This sample dataset contains:
|
59 |
+
- 5 Japanese natural language queries describing programming tasks
|
60 |
+
- 5 corresponding Python code snippets
|
61 |
+
- 5 relevance judgments connecting queries to code
|
62 |
+
|
63 |
+
The data has been slightly modified for demonstration purposes.
|
64 |
+
|
65 |
+
## License
|
66 |
+
|
67 |
+
Please refer to the CC BY-SA 4.0 license terms.
|
corpus/corpus-00000-of-00001.parquet
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bbc9a9bab4977312f8f2535159ebf1bccda79d7137626580addaa31a3eb87eb2
|
3 |
+
size 2442
|
create_sample_dataset.py
ADDED
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
|
3 |
+
import pandas as pd
|
4 |
+
import random
|
5 |
+
from pathlib import Path
|
6 |
+
|
7 |
+
def create_sample_dataset():
|
8 |
+
# Load original data
|
9 |
+
base_path = Path("/Users/fodizoltan/Projects/toptal/voyageai/tmp/rteb/data/JapaneseCode1Retrieval")
|
10 |
+
|
11 |
+
corpus_df = pd.read_parquet(base_path / "corpus" / "corpus-00000-of-00001.parquet")
|
12 |
+
queries_df = pd.read_parquet(base_path / "queries" / "queries-00000-of-00001.parquet")
|
13 |
+
qrels_df = pd.read_parquet(base_path / "data" / "test-00000-of-00001.parquet")
|
14 |
+
|
15 |
+
print(f"Loaded {len(corpus_df)} corpus docs, {len(queries_df)} queries, {len(qrels_df)} qrels")
|
16 |
+
|
17 |
+
# Get first 5 queries
|
18 |
+
first_5_queries = queries_df.head(5).copy()
|
19 |
+
print("First 5 queries:")
|
20 |
+
for _, row in first_5_queries.iterrows():
|
21 |
+
print(f" {row['_id']}: {row['text']}")
|
22 |
+
|
23 |
+
# Get related corpus documents from qrels
|
24 |
+
first_5_query_ids = first_5_queries['_id'].tolist()
|
25 |
+
related_qrels = qrels_df[qrels_df['query-id'].isin(first_5_query_ids)].copy()
|
26 |
+
related_corpus_ids = related_qrels['corpus-id'].unique()
|
27 |
+
related_corpus = corpus_df[corpus_df['_id'].isin(related_corpus_ids)].copy()
|
28 |
+
|
29 |
+
print(f"Found {len(related_corpus)} related corpus documents")
|
30 |
+
print(f"Found {len(related_qrels)} qrels")
|
31 |
+
|
32 |
+
# Tweak the data
|
33 |
+
tweaked_queries = tweak_queries(first_5_queries)
|
34 |
+
tweaked_corpus = tweak_corpus(related_corpus)
|
35 |
+
|
36 |
+
return tweaked_corpus, tweaked_queries, related_qrels
|
37 |
+
|
38 |
+
def tweak_queries(queries_df):
|
39 |
+
"""Tweak Japanese queries by changing a few words/characters"""
|
40 |
+
tweaked = queries_df.copy()
|
41 |
+
|
42 |
+
# Simple tweaks for Japanese text
|
43 |
+
tweaks = {
|
44 |
+
'リスト': 'リスト配列',
|
45 |
+
'要素': 'エレメント',
|
46 |
+
'整数': '数値',
|
47 |
+
'文字列': 'ストリング',
|
48 |
+
'変数': 'パラメータ',
|
49 |
+
'する': 'を行う',
|
50 |
+
'の': 'が持つ'
|
51 |
+
}
|
52 |
+
|
53 |
+
for idx, row in tweaked.iterrows():
|
54 |
+
text = row['text']
|
55 |
+
# Apply random tweaks
|
56 |
+
for old, new in tweaks.items():
|
57 |
+
if old in text and random.random() < 0.3: # 30% chance to apply each tweak
|
58 |
+
text = text.replace(old, new, 1) # Replace only first occurrence
|
59 |
+
tweaked.at[idx, 'text'] = text
|
60 |
+
|
61 |
+
return tweaked
|
62 |
+
|
63 |
+
def tweak_corpus(corpus_df):
|
64 |
+
"""Tweak Python code by changing variable names and some function calls"""
|
65 |
+
tweaked = corpus_df.copy()
|
66 |
+
|
67 |
+
# Simple code tweaks
|
68 |
+
code_tweaks = {
|
69 |
+
'x': 'data',
|
70 |
+
'i': 'idx',
|
71 |
+
'd': 'digit',
|
72 |
+
'enumerate': 'enumerate_items',
|
73 |
+
'sum': 'total',
|
74 |
+
'len': 'length',
|
75 |
+
'str': 'string',
|
76 |
+
'int': 'integer',
|
77 |
+
'list': 'array'
|
78 |
+
}
|
79 |
+
|
80 |
+
for idx, row in tweaked.iterrows():
|
81 |
+
text = row['text']
|
82 |
+
# Apply random tweaks
|
83 |
+
for old, new in code_tweaks.items():
|
84 |
+
if old in text and random.random() < 0.4: # 40% chance to apply each tweak
|
85 |
+
# Only replace whole words to avoid breaking code
|
86 |
+
import re
|
87 |
+
pattern = r'\b' + re.escape(old) + r'\b'
|
88 |
+
if re.search(pattern, text):
|
89 |
+
text = re.sub(pattern, new, text, count=1)
|
90 |
+
tweaked.at[idx, 'text'] = text
|
91 |
+
|
92 |
+
return tweaked
|
93 |
+
|
94 |
+
def save_sample_dataset(corpus_df, queries_df, qrels_df):
|
95 |
+
"""Save sample dataset in MTEB format"""
|
96 |
+
sample_path = Path("/Users/fodizoltan/Projects/toptal/voyageai/tmp/rteb/data/JapaneseCode1Retrieval-sample")
|
97 |
+
sample_path.mkdir(exist_ok=True)
|
98 |
+
|
99 |
+
# Create subdirectories
|
100 |
+
(sample_path / "corpus").mkdir(exist_ok=True)
|
101 |
+
(sample_path / "queries").mkdir(exist_ok=True)
|
102 |
+
(sample_path / "data").mkdir(exist_ok=True)
|
103 |
+
|
104 |
+
# Save parquet files
|
105 |
+
corpus_df.to_parquet(sample_path / "corpus" / "corpus-00000-of-00001.parquet", index=False)
|
106 |
+
queries_df.to_parquet(sample_path / "queries" / "queries-00000-of-00001.parquet", index=False)
|
107 |
+
qrels_df.to_parquet(sample_path / "data" / "test-00000-of-00001.parquet", index=False)
|
108 |
+
|
109 |
+
print(f"Sample dataset saved to {sample_path}")
|
110 |
+
print(f" - Corpus: {len(corpus_df)} documents")
|
111 |
+
print(f" - Queries: {len(queries_df)} queries")
|
112 |
+
print(f" - Qrels: {len(qrels_df)} relevance judgments")
|
113 |
+
|
114 |
+
if __name__ == "__main__":
|
115 |
+
corpus, queries, qrels = create_sample_dataset()
|
116 |
+
save_sample_dataset(corpus, queries, qrels)
|
data/test-00000-of-00001.parquet
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5442e12e4ed25507ee3a90e03076eb79e67ae8287033e1bf798b8de3733f7871
|
3 |
+
size 2256
|
queries/queries-00000-of-00001.parquet
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:25661ad4a4d012f953ba8fb1f013af86c8ac4252c25aa16e16770e22f970039d
|
3 |
+
size 2211
|