-
Reorganize: drop -boxed/-buggy/-fixed variants, add metadata column for slime030 dispatch
Code data (4 files, schema: data_source/prompt/ability/reward_model/extra_info/metadata):
train.parquet (24287) — primeintellect + taco + livecodebench
test_livecodebench.parquet (279)
test_codeforces.parquet (408)
test_humanevalplus.parquet (163)
Math data (7 files, all \boxed{} prompts):
dapo-math-17k.parquet (17395)
test_aime2024.parquet (30)
test_aime2025.parquet (30)
test_amc23.parquet (40)
test_math500.parquet (500)
test_minerva.parquet (272)
test_olympiadbench.parquet (674)
New 'metadata' column carries {data_source, ability, **extra_info} so slime030's
custom_rm dispatcher routes code samples to the correct rllm scorer
(lcb / primeintellect / humanevalplus / etc.) by data_source instead of
ground_truth format heuristics. All string fields use pyarrow large_string for
pandas compatibility with nested chunked arrays.
