Overview
Research Questions
Q1
最小データ量: 新タスク FT で目標性能を達成する最小 ep 数
min data 4 段階 (10/30/50/100 ep) × success rate
Q2
限界 iteration: catastrophic forgetting の限界
sorting OOD color generalization (赤・青 学習 → 緑・紫 評価)
dataset: ファインチューニングに使う… See the full description on the dataset page:
https://huggingface.co/datasets/Harumo/so101-ft-adaptation-study.