Implement on-policy distillation: have the student generate completions, score them with a teacher, and…
Implement on-policy distillation: have the student generate completions, score them with a teacher, and…: a task in traj2envs-rl: RL tasks generated with traj2envs (Harbor dataset). Scenario parameters that must reproduce those winners: - Multimodal: length 8; teacher sticky(8, start prob one=0.5…
The task
Scenario parameters that must reproduce those winners: - Multimodal: length 8; teacher `sticky(8, start_prob_one=0.5, prob_one_after_zero=0.02, prob_one_after_one=0.98)`; initial student `independent(8, 0.5)`; offline trains on `make_top_corpus(teacher, 1)` for 500 steps at lr 0.05; on-policy trains 500 steps, 128…
Part of harness-hub/traj2envs-rl.