Trained by reinforcement learning on rows of eight OpenMathReasoning GenSelect candidates with one to seven of them correct, does Qwen3.5-9B's accuracy with eight candidates on held-out rows of the same distribution rise above its untrained accuracy and above the majority vote of the eight? - W&B runs, all in pqin/atlas-grpo: the four-card-plus-three memory check attempts nrtc229i, pkn4gkpq and tzdiu97r (the last one passed), and the eight-card sync memory check nhwjntqg (passed). - README.md: this front page. - data/omr/merged.jsonl (884 MB): one line per problem of the GenSelect split whose answer was extracted from the forum, the candidates merged across the problem's prompts and…
Independent publisher
Peijia Qin
t2ance
Models in Library0
Datasets in Library3
Models on Hugging Face94
Followers1
Compare the published evidence
This profile connects the 3 model and dataset records shown below to their upstream cards and sources. The library counts describe SAVRN's coverage, rather than this publisher's entire output.
Compare model specifications, source dates, and hardware estimates on the individual records. Publisher identity verification is separate from a model's quality, deployment readiness, or suitability for your work.
Datasets
Read this first. The reading copy of this directory is t2ance/atlas-experiments under 31-strengthening-candidate-verification-under-rl/; the saved training steps and the per-token training arrays are on the Hugging Face repository t2ance/atlas-31-strengthening-candidate-verification-under-rl only. How can reinforcement learning make the orchestrator's comparing and verifying of candidates emerge or grow stronger, so that accuracy rises with the candidates read toward the best of them and the policy settles neither on one call nor on the vote? - W&B runs, in pqin/atlas-grpo: the training r31-techniques-grpo, started 2026-09-20 22:39 UTC (group r31-techniques-grpo: the trainer's run 1b5e1ae4…
Read this first. The reading copy of this directory is t2ance/atlas-experiments under 32-turn-level-actor-critic/; the saved training steps and the per-token training arrays are only in the Hugging Face repository t2ance/atlas-32-turn-level-actor-critic. Does a critic that predicts the return at the start of each turn, and is supervised there alone, learn on the orchestrator's rollouts (report 30's token-level critic did not), and does a policy trained on its turn-level advantages keep reading candidates where reading pays (report 31's critic-free training dropped to 0.93 explores a problem with the accuracy of its submissions unchanged)? - W&B runs, in pqin/atlas-grpo: the smoke…