# Supervised Finetuning

Researchers curated OpenTumorBoard from 219 recorded tumor board meetings totaling 12,534 minutes of YouTube video, yielding 611 patient cases, 19,157 discussion turns and ten specialist roles, and used it for two tasks—single Specialist Turn responses and full Board Simulation—finding that across 14 general-purpose and medical LLMs the best model reached only 3.43/5 in clinical equivalence and 2.78/5 in alignment with true board conclusions, while supervised finetuning plus reinforcement learning on Qwen2.5-VL-3B raised conclusion alignment from 1.58 to 1.86.

Researchers curated OpenTumorBoard from 219 recorded tumor board meetings totaling 12,534 minutes of YouTube video, yielding 611 patient cases, 19,157 discussion turns and ten specialist roles, and used it for two tasks—single Specialist Turn responses and full Board Simulation—finding that across 14 general-purpose and medical LLMs the best model reached only 3.43/5 in clinical equivalence and 2.78/5 in alignment with true board conclusions, while supervised finetuning plus reinforcement learning on Qwen2.5-VL-3B raised conclusion alignment from 1.58 to 1.86.

SkillGym proposes an automatic pipeline that crawls skills from the internet and keeps those whose workflows run reproducibly offline, uses a builder-reviewer agent system to construct skill-critical tasks across four reasoning structures each with a reference solution and an executable verifier, then collects successful trajectories from three teacher models and four agent harnesses for supervised finetuning, improving six models from 2B to 122B parameters on 22 of 24 comparisons across four skill-use benchmarks, with the 9B model outperforming Qwen3.5-397B-A17B on the SkillGym test set and SkillEval and raising the rate of reading the relevant skill from 28% to 96%.

SkillGym proposes an automatic pipeline that crawls skills from the internet and keeps those whose workflows run reproducibly offline, uses a builder-reviewer agent system to construct skill-critical tasks across four reasoning structures each with a reference solution and an executable verifier, then collects successful trajectories from three teacher models and four agent harnesses for supervised finetuning, improving six models from 2B to 122B parameters on 22 of 24 comparisons across four skill-use benchmarks, with the 9B model outperforming Qwen3.5-397B-A17B on the SkillGym test set and SkillEval and raising the rate of reading the relevant skill from 28% to 96%.