Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
230
2k
End of preview. Expand in Data Studio

RoboDojo stack_bowls: on-policy self-distillation of lessons

Data for distilling written lessons ("skills") into a robot agent's weights. A planning model controls two simulated robot arms through tools to stack three bowls (RoboDojo stack_bowls, ARX X5 arms, Isaac Sim). The teacher is the model with a library of lessons pinned in its system prompt; the student is the same model without them. Each turn holds a completion sampled from the student and the teacher's top-64 next-token distribution at every position of that completion, so on-policy distillation (per-token reverse KL to the teacher) can be reproduced without running the teacher.

  • Model: Qwen3.8-Flash-Next (served NVFP4 with vLLM for sampling; BF16 for the teacher pass), reasoning on.
  • Episodes: 24 recorded agent episodes on training layouts 1-4 (episodes.jsonl; groups name the lesson condition of the recorded run: r1_nolessons, r1_lessons, r1_distilled, r2, r3, r4; each group has one episode per layout). native_score is null where the recorded episode was cut off before scoring.
  • Turns: 4,736 (one per model call of the recorded episodes; context-compaction calls excluded).
  • Lessons: the teacher's pinned lessons are in lessons/teacher_r4/ (all-tasks.md, stack_bowls.md).

Layout

episodes.jsonl                     one row per episode: run, group, layout, turns, images, native_score, stop_reason, steps
lessons/teacher_r4/*.md            lessons pinned in the teacher's system prompt
data/<run>/turns.jsonl.gz          one JSON object per turn (below)
data/<run>/images/*.webp           camera images referenced by the messages
data/<run>/targets/<exchange>.npz  exact OPSD targets for the turn

Each line of turns.jsonl.gz:

field content
run, exchange, group, layout episode id, model-call index within the episode, lesson condition, layout index
student_messages OpenAI-style chat messages of the student context (no lessons). Image parts hold a path relative to data/<run>/ in image_url.url.
teacher_messages the same history with the lessons pinned in the system prompt and lesson-file reads answered from the lesson library
tools tool schemas offered to the model
request_params remaining request fields (model, reasoning effort, tool choice, ...)
student_sample the student's sampled reasoning, content, tool_calls, finish_reason, num_completion_tokens
targets path of the turn's .npz

Each .npz: student_prompt_ids and teacher_prompt_ids (prompt token ids as tokenized by vLLM's chat template, images expanded), completion_ids (the student sample), teacher_topk_ids and teacher_topk_logprobs ([n, 64], teacher top-64 at each completion position; training used at most the first 8,192 completion tokens).

The episodes' tool histories are the recorded ones; only the final completion of each turn is a fresh student sample, so the turn's prompt is off-policy for the student and the completion is on-policy.

Notes

  • Machine-specific paths are replaced by <run_dir>, <runs_root> and ~.
  • native_score is the simulator's task score (1.0 = all three bowls nested and arms returned).
  • The teacher's lessons were written from reading earlier episodes' traces; see lessons/teacher_r4.
Downloads last month
163