---
license: mit
task_categories:
- reinforcement-learning
tags:
- sokoban
- trajectory
- teacher-student
- llm-agent
configs:
- config_name: default
  data_files:
  - split: teacher
    path: data/teacher.parquet
  - split: student
    path: data/student.parquet
  - split: student_prefix
    path: data/student_prefix.parquet
  - split: teacher_prefix
    path: data/teacher_prefix.parquet
---

# Sokoban-Trajectories

Rollout trajectories for the **Sokoban** (推箱子) environment generated using the
[RAGEN](https://github.com/RAGEN-AI/RAGEN) framework.

## Models

| Role | Model |
|------|-------|
| Teacher | Qwen/Qwen2.5-14B-Instruct |
| Student | Qwen/Qwen2.5-3B-Instruct |

## Environment Settings

| Setting | Value |
|---------|-------|
| Room size | 6×6 |
| Number of boxes | 1 |
| Max turns per episode | 10 |
| **Actions per turn** | **up to 2** (`max_actions_per_turn=2`) |
| Max actions per trajectory | 20 (10 turns × 2 actions) |

> **Note:** 1 turn = up to 2 actions. The LLM is called once per turn and may output
> 1 or 2 actions. All turn-based statistics below count LLM calls, not individual actions.

## Data Scale

- 500 problems × 4 trajectories each = **2000 trajectories per split**
- Seeds: val base seed 123 (problems 123–622)

## Cutoff

Teacher average turns = **4.21** → cutoff = `floor(4.21 / 2)` = **2 turns** (≤4 actions)

## Splits

| Split | Description |
|-------|-------------|
| `teacher` | Full rollouts by teacher (14B) from start to finish |
| `student` | Full rollouts by student (3B) from start to finish |
| `student_prefix` | **Step 4**: Student runs first **2 turns**, teacher completes the rest |
| `teacher_prefix` | **Step 5**: Teacher runs first **2 turns**, student completes the rest |

## Results

| Setting | success | pass@4 | avg turns |
|---------|---------|--------|-----------|
| Student only | 0.066 | 0.160 | 9.31 |
| Step 4: student→teacher | 0.487 | 0.682 | 7.47 |
| Step 5: teacher→student | 0.223 | 0.402 | 8.30 |
| Teacher only | 0.655 | 0.844 | 4.21 |

Key finding: teacher completion after student prefix (Step 4) substantially
improves over student-only, but the student struggles to finish after a teacher
prefix (Step 5), indicating the student's ability to complete partially-solved
puzzles is the bottleneck.

## Schema

Each row is one trajectory:

| Column | Type | Description |
|--------|------|-------------|
| `env_id` | int | Unique environment index (0–1999) |
| `group_id` | int | Problem group index (0–499); 4 trajectories share the same problem |
| `turn_count` | int | Number of LLM turns taken (1 turn = up to 2 actions) |
| `messages` | list[dict] | Full conversation: `[{role, content}, ...]` |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("CL-From-Nothing/Sokoban-Trajectories")

# Full teacher rollouts
teacher = ds["teacher"]

# Step 4: student prefix (2 turns) + teacher completion
step4 = ds["student_prefix"]

# Access messages for first trajectory
print(step4[0]["messages"])
```