| --- |
| license: apache-2.0 |
| task_categories: |
| - text-generation |
| tags: |
| - agent |
| - scienceworld |
| - distillation |
| - rose |
| language: |
| - en |
| --- |
| |
| # student_partial_sciworld |
|
|
| Partial SciWorld trajectories: **Qwen3-1.7B (base) acting for its first 5 turns** on |
| the 2,120-task training split. These are the *student prefixes* an online-ROSE run sees |
| before the teacher takes over — captured separately so the take-over point can be |
| studied, or teacher continuations generated offline. |
|
|
| ## What is here |
|
|
| | | | |
| |---|---| |
| | rows | **2,120** (one per training task) | |
| | file | `student_partial_sciworld.jsonl` (11.4 MB) | |
| | model | `Qwen/Qwen3-1.7B`, untrained base — the starting point of our ROSE/OPD runs | |
| | turns | 5 (mean 4.96; a few episodes end early) | |
| | sampling | temperature 1.0, `max_new_tokens` 512, context 16384 | |
| | mean score | **0.0352** | |
|
|
| Sampling matches what the student does inside the online-ROSE loop, so these prefixes |
| are distributionally the same as the ones the teacher continues during training. |
|
|
| ## Schema |
|
|
| ```json |
| { |
| "item_id": "sciworld_0", |
| "task": "boil", |
| "var": 0, |
| "score": 0.0, // ScienceWorld completion in [0,1] after 5 turns |
| "success": 0, // score == 1 |
| "rounds": 5, |
| "conversations": [ // system + 24-action instruction + ack, then obs/action pairs |
| {"role": "user", "content": "You are an agent for science world. ..."}, |
| {"role": "assistant", "content": "OK. I'll follow your instructions ..."}, |
| {"role": "user", "content": "<observation>"}, |
| {"role": "assistant", "content": "Thought: ...\nAction: ..."} |
| ] |
| } |
| ``` |
|
|
| `conversations` is the full AgentGym-protocol dialogue, ready to re-render with any |
| chat template. |
|
|
| ## Why 5 turns |
|
|
| Five turns is where online ROSE cuts. At that point the student has explored but has |
| essentially completed nothing — mean score **0.0352**, versus **0.0479** when the same |
| base model runs the full 20 turns. The take-over lands after the student has committed |
| to an approach but before the episode is decided. |
|
|
| For reference, on the same environment (200 held-out tasks, ≤20 turns, greedy): |
|
|
| | | Score | Success | Pass(>0) | |
| |---|---|---|---| |
| | Qwen3-1.7B base | 0.0479 | 0.0000 | 0.1650 | |
| | SFT on Qwen3-32B trajectories | 0.2077 | 0.0450 | 0.4100 | |
| | online ROSE (5 student + 5 teacher) | 0.3127 | 0.0600 | 0.6600 | |
| | Qwen3-32B teacher | 0.4346 | 0.1100 | 0.9350 | |
|
|
| ## Environment |
|
|
| ScienceWorld via the AgentGym protocol. Tasks are `(task_type, variation)` pairs with |
| seven variation families excluded (`5-1 5-2 9-1 9-2 9-3 10-1 10-2`), giving 4,639 games |
| total; this split holds 2,120 of them. The prompt is the standard 24-action REACT |
| instruction, and actions are parsed from the first non-empty line after `Action:`. |
|
|
| Scores are ScienceWorld's own completion signal, not a learned reward. |
|
|
