metadata
license: mit
DeepScaleR-Easy-10k
RL training set of 10,000 DeepScaleR questions that Gemma-3-4B-IT solved 4/4 (temp 1.0, 4 samples, LENIENT math_verify grading), split train 9,500 / val 500.
Creation (seed 42 throughout):
- Gemma-3-4B-IT generated 4 responses per DeepScaleR question (40,315 rows).
- Dedup v2: questions grouped under max-strength normalization (case,
whitespace, prose punctuation, LaTeX formatting); duplicate groups whose gold
answers are not all math_verify-equivalent were removed entirely (167 groups /
567 rows — label conflicts); one random copy kept per surviving group
-> 38,796 unique questions. Script:
rl-distill-scripts/data/dedup_deepscaler_it_gen.py. - 10,000 sampled uniformly from the lenient 4/4 bucket (10,053 available).
- 500 sampled uniformly as validation; remaining 9,500 are train.
Files: *_train.parquet and *_val500.parquet (500 unique held-out questions).
For mean@16 validation, replicate the val rows 16x locally (verl samples one
response per row). verl format: data_source "math" (routes to the repo's
math_verify scorer), prompt = single user message, reward_model.ground_truth,
extra_info carries lenient/strict pass counts.
Built by rl-distill-scripts/data/build_deepscaler_easy_medium.py in
JasonWei05/rl-distill.