2026-10-21T17:09:37.351706
FrontierVEP comparison. Base + RL: HuggingFaceBio/Qwen3-14B-AGRPO v00.01, through checkpoint 190. Mid-training + RL: HuggingFaceBio/Qwen3-14B-Instruct-Mid-300B-AGRPO v15.01, through checkpoint 200. Raw vepqa_accuracy_reward from final trainer_state.json with trailing 10-step means. Evaluation points: latest successful vepqa_single_1k primary scores every 10 steps from HuggingFaceBio/post-training-evals.
image/svg+xml
Matplotlib v3.10.8, https://matplotlib.org/