PugarHuda commited on
Commit
28f45ae
·
verified ·
1 Parent(s): 7debf49

Benak release

Browse files
Files changed (4) hide show
  1. README.md +1 -1
  2. code/build.py +1 -1
  3. data/sea_thinking.jsonl +0 -0
  4. results.json +1 -1
README.md CHANGED
@@ -57,7 +57,7 @@ configs:
57
  |---|---|---|
58
  | `atlas_runs` | 3,200 | Belebele, 19 languages × 40 parallel items × 4 prompting conditions (+ effort=high subset). Every call is included, with full reasoning, answer, grade, token usage and GlotLID profile. |
59
  | `math_runs` | 1,276 | 58 code-verified math word problems × 9 languages × {C0, C2}, plus C3 (the `benak_lang` wrapper) for Indonesian, Javanese, Sundanese and Thai. |
60
- | `sea_thinking` | 1,453 | **The fix.** Correct answers whose reasoning is ≥ 80% in the target language, in the [HuggingFaceH4/Multilingual-Thinking](https://huggingface.co/datasets/HuggingFaceH4/Multilingual-Thinking) schema (`reasoning_language, developer, user, analysis, final, messages`). That is the format Reflection's own TRL recipe uses, so it can be used directly for SFT. `source` marks native traces vs Beam-rewritten traces. |
61
  | `replication_profiled` | 4,139 | Beam traces that other entrants had already published (neulketing Korean, esscor TamilBench), re-scored with our detector. These cost $0 of inference. |
62
  | `grammar_book_bugis.json` | 20 calls | Can a book in Beam's 1M-token context teach it Buginese? Includes hypotheses, chrF++ scores and the reasoning-language probe. |
63
 
 
57
  |---|---|---|
58
  | `atlas_runs` | 3,200 | Belebele, 19 languages × 40 parallel items × 4 prompting conditions (+ effort=high subset). Every call is included, with full reasoning, answer, grade, token usage and GlotLID profile. |
59
  | `math_runs` | 1,276 | 58 code-verified math word problems × 9 languages × {C0, C2}, plus C3 (the `benak_lang` wrapper) for Indonesian, Javanese, Sundanese and Thai. |
60
+ | `sea_thinking` | 1,204 | **The fix.** Correct answers whose reasoning is ≥ 80% in the target language (18 non-English languages; most rows are Southeast Asian), in the [HuggingFaceH4/Multilingual-Thinking](https://huggingface.co/datasets/HuggingFaceH4/Multilingual-Thinking) schema (`reasoning_language, developer, user, analysis, final, messages`). That is the format Reflection's own TRL recipe uses, so it can be used directly for SFT. `source` marks native traces vs Beam-rewritten traces. |
61
  | `replication_profiled` | 4,139 | Beam traces that other entrants had already published (neulketing Korean, esscor TamilBench), re-scored with our detector. These cost $0 of inference. |
62
  | `grammar_book_bugis.json` | 20 calls | Can a book in Beam's 1M-token context teach it Buginese? Includes hypotheses, chrF++ scores and the reasoning-language probe. |
63
 
code/build.py CHANGED
@@ -35,7 +35,7 @@ think = []
35
  _bel = {}
36
  for r in math_ + atlas + rew:
37
  ts = (r.get("lid") or {}).get("target_share")
38
- if r.get("correct") and ts is not None and ts >= 0.8 and r.get("content"):
39
  lang = r["lang"]
40
  if "link" in r: # Belebele item: rebuild the exact prompt Beam saw (C0-C3 share the same user turn)
41
  from atlas import mc_prompt
 
35
  _bel = {}
36
  for r in math_ + atlas + rew:
37
  ts = (r.get("lid") or {}).get("target_share")
38
+ if r.get("correct") and ts is not None and ts >= 0.8 and r.get("content") and r["lang"] != "eng_Latn":
39
  lang = r["lang"]
40
  if "link" in r: # Belebele item: rebuild the exact prompt Beam saw (C0-C3 share the same user turn)
41
  from atlas import mc_prompt
data/sea_thinking.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
results.json CHANGED
@@ -2814,5 +2814,5 @@
2814
  "huggingface": 2116
2815
  },
2816
  "cost_usd": 2.8987,
2817
- "sea_thinking_rows": 1453
2818
  }
 
2814
  "huggingface": 2116
2815
  },
2816
  "cost_usd": 2.8987,
2817
+ "sea_thinking_rows": 1204
2818
  }