license: mit
DeepScaleR-Medium-20k
RL training set of 20,000 DeepScaleR questions spanning difficulty (Gemma-3-4B-IT pass rate at temp 1.0, 4 samples, LENIENT math_verify grading): the same 10,000 4/4 questions as DeepScaleR-Easy-10k + 3,000 from 3/4 + 3,000 from 2/4 + 4,000 from 1/4. 500 random questions (across all difficulties) held out as validation; train 19,500.
Creation (seed 42 throughout): identical dedup-v2 pipeline as DeepScaleR-Easy-10k (see that README): max-strength question dedup, conflicting-gold groups removed via math_verify equivalence, one random copy per group -> 38,796 unique questions; buckets sampled uniformly without replacement.
NOTE: because the 10k easy questions are shared, DeepScaleR-Easy-10k's val questions may appear in this set's TRAIN split (and vice versa). Do not evaluate a Medium-trained model on Easy-10k's val (or vice versa); each set is self-consistent only with its own split.
Files: *_train.parquet and *_val500.parquet (500 unique held-out questions).
For mean@16 validation, replicate the val rows 16x locally. verl format:
data_source "math", 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.