FrozenLake-Trajectories

Rollout trajectories for the FrozenLake environment (4×4 grid, is_slippery=True) generated using the RAGEN framework.

Models

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

Environment

• Grid: 4×4, stochastic (is_slippery=True)
• 500 problems × 4 trajectories each = 2000 trajectories per split
• Max turns: 10 (up to 2 actions per turn)
• Seeds: val base seed 123 (problems 123–622)

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 1 turn, teacher completes the rest
teacher_prefix	Step 5: Teacher runs first 1 turn, student completes the rest

Cutoff

Teacher average turns = 2.04 → cutoff = floor(2.04 / 2) = 1 turn

Results

Setting	success	pass@4	avg num_actions
Student only	0.241	0.482	6.37
Step 4: student→teacher	0.722	0.908	4.09
Step 5: teacher→student	~0.780	0.784	4.59
Teacher only	~1.000	1.000	2.04

Key finding: teacher completion (Step 4) recovers most of the performance gap, while student completion after teacher prefix (Step 5) is less effective, suggesting the student's "finishing ability" 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
messages	list[dict]	Full conversation: [{role, content}, ...]

Usage

from datasets import load_dataset

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

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

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

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