JWei05/DeepScaleR-Easy-Medium-Hard-Gemma-26B-PT-10k

DeepScaleR Easy/Medium/Hard — Gemma 4 26B-A4B PT

This dataset contains 9,900 unique, deduplicated DeepScaleR math questions for reinforcement-learning experiments. Difficulty is defined by how often the pretrained google/gemma-4-26B-A4B teacher solved each question across eight temperature-1 samples under the same rule-based grader used by the RL training pipeline.

The Hub dataset has three configurations—easy, medium, and hard—and each configuration has a train split with 3,000 questions and a disjoint validation split with 300 questions.

Configurations and splits

Configuration	Teacher score band	Available before sampling	Train	Validation	Total selected
easy	4/8-8/8	3,314	3,000	300	3,300
medium	2/8-3/8	3,498	3,000	300	3,300
hard	1/8	5,342	3,000	300	3,300

The 0/8 bucket is intentionally excluded. In this dataset, hard means exactly 1/8, matching the band definition used for the associated E2B/E4B experiments.

Exact score composition after sampling

Configuration	Split	Number of rows by teacher score
easy	train	4/8: 838, 5/8: 687, 6/8: 616, 7/8: 544, 8/8: 315
easy	validation	4/8: 96, 5/8: 64, 6/8: 64, 7/8: 52, 8/8: 24
medium	train	2/8: 1,896, 3/8: 1,104
medium	validation	2/8: 192, 3/8: 108
hard	train	1/8: 3,000
hard	validation	1/8: 300

Construction

1. Start from agentica-org/DeepScaleR-Preview-Dataset (40,315 original rows, MIT license).
2. Use the existing dedup-v2 view with 38,796 unique questions. That pipeline groups questions under strong text/LaTeX normalization, removes duplicate groups with conflicting non-equivalent gold answers, and keeps one deterministic representative from each surviving group. The exact local source Parquet has SHA-256 4de0a409372f3af6cb30fc47cd9e12f98114e4a8d51664fefd87a08be4c62b5a.
3. Grade eight independently sampled teacher responses per question using the current RL-training grading contract described below. The exact question-level grade Parquet has SHA-256 e0945e65ad3f176541c84b0a62094078a2f64a6a7607de082c495407db812d31.
4. Form disjoint score bands: hard = 1/8, medium = 2–3/8, easy = 4–8/8.
5. Within each band, rank every UID by SHA256("selection\042\0<band>\0<uid>") and take the lowest 3,300 ranks without replacement.
6. Within each selected set, rank UIDs by SHA256("validation\043\0<band>\0<uid>"); the lowest 300 become validation and the remaining 3,000 become train.

Stable hash ranking makes the result independent of pandas, NumPy, and Python random-number-generator versions. manifest.json contains the output file hashes, UID-set hashes, source hashes, score counts, and all construction settings.

Teacher generation protocol

• Model: google/gemma-4-26B-A4B (pretrained/base), revision 24548b62aa021d562695c04aaf7758a1ea47990b.
• Eight responses per question.
• Sampling: temperature 1, top-p 1, top-k -1.
• Maximum response length: 4,096 tokens.
• Maximum model context: 9,216 tokens.
• Prompt: the unified 12-shot math template gemma3_it_fewshot_math.jinja, SHA-256 27b8801d8b61a413a9bb3b54b6f55e16217eff3e55f7c560377c8a162dd63c1c.
• Generation stops: <end_of_turn> and <start_of_turn>.

The data rows store the plain problem as the single user message. The 12-shot template is applied by the RL training configuration at tokenization time; it is not duplicated into every dataset row.

RL grading contract

For each response, the grader extracts the last well-formed \boxed{...} or \fbox{...} answer. A response is correct if either Hugging Face math_verify or the bounded Miles-style SymPy fallback accepts equivalence with the gold answer.

• math_verify timeout: 30 seconds.
• Miles-SymPy timeout: 5 seconds.
• Evaluation-time math_verify.py SHA-256: e403de40ac42632fcd703763983685f0cca23af4a8e0e4c334e5ff0cb9236db2.
• Evaluation-time miles_sympy.py SHA-256: 8fe9cd48189bd03ce667ac24193e9f3394c22defe78076e5ff5244c831628795.

This is the current RL grader, not the earlier whole-output lenient grader or the earlier exactly-one-box strict grader.

Schema

Each row is directly usable by verl-style rule-reward math training:

• uid: stable UID inherited from the deduplicated source.
• data_source: always math.
• prompt: one plain user message containing the problem.
• reward_model: {"style": "rule", "ground_truth": <gold answer>}.
• extra_info: split/band provenance, teacher model and revision, all eight per-trace correctness flags, teacher pass count/rate, prompt and sampler settings, grader contract, deterministic ranks, and sampling seeds.

Reproduction

From the rl-distill repository, using its uv environment:

.venv/bin/python rl-distill-scripts/data/build_deepscaler_gemma4_26b_difficulty_dataset.py

The default build writes the six Parquet files, this README, and manifest.json locally. Pass --upload to publish the same files to JWei05/DeepScaleR-Easy-Medium-Hard-Gemma-26B-PT-10k.

Intended use and limitations

These labels measure empirical solvability by one specific pretrained teacher, prompt, sampler, response limit, and grader. They are useful curriculum bands, not universal or human difficulty labels. The easy candidate pool contains only 3,314 questions, so its 3,300-question sample retains nearly the entire available easy band. Validation splits are disjoint from training within and across configurations, but they are not intended as a replacement for established external math benchmarks.