datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
maxrl_qwen3_4B_base_polaris_rollouts
MaxRL Qwen3-4B-Base training rollouts (POLARIS math prompts)
Every training rollout from an online RL run, with exact token ids, sampling
log-probs, and raw rewards — usable as a replay buffer to study off-policy RL
for LLM reasoning completely offline.
The run: Qwen3-4B-Base trained with the maxRL advantage estimator
(A = (r - mean)/(mean + eps), group mean over 16 rollouts per prompt;
maxRL paper) and a pure REINFORCE loss
(L = -A * log pi; no importance ratio, no clipping, no… See the full description on the dataset page: https://huggingface.co/datasets/ftajwar/maxrl_qwen3_4B_base_polaris_rollouts.polaris-53k-repaired
POLARIS-53K, label-repaired
49,289 of the 53,291 rows in
POLARIS-Project/Polaris-Dataset-53K,
with 4,580 stored answers corrected and 4,002 rows removed as unrepairable.
Measurements on the source set put its bad-label rate at roughly 15.9%
[14.3, 17.6] (two independent detectors agreeing on a 2,000-row sample).
Mislabelled rows are not uniformly distributed: they concentrate in the problems
models fail, which is exactly where a difficulty-calibration pipeline looks.… See the full description on the dataset page: https://huggingface.co/datasets/joanvelja/polaris-53k-repaired.Polaris-hard-w-solutions-24209
Polaris-Hard-w-Solutions
24,209 hard competition-math problems (the hardest difficulty bands of the
Polaris dataset) paired with
two verified solutions each: a full original solution and a concise summarized solution. Every
retained problem has a machine-verifiable final answer, every solution's boxed answer grades correct
against the reference (sympy-based grading), and the summarized solutions have additionally been put
through a reasoning-rigor pass (see step 5 below).
This… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/Polaris-hard-w-solutions-24209.polaris_rose_rollouts_olmo3-7b_from_qwen3-30b-a3b_cutoff4096_240steps
Cross-tokenizer ROSE rollouts — Olmo-3-7B-Think-SFT ← Qwen3-30B-A3B-Thinking-2507
Every assembled row of a complete 240-step online-ROSE run: 61,440 rows, the teacher's
actual continuation for each, and the token accounting behind it.
The student writes a 4096-token prefix in its own vocabulary (100278). That prefix is
decoded to text, the teacher is shown it under its own chat template, and the teacher's
reply comes back as text and is tokenised into the student's vocabulary.… See the full description on the dataset page: https://huggingface.co/datasets/SeanWang0027/polaris_rose_rollouts_olmo3-7b_from_qwen3-30b-a3b_cutoff4096_240steps.