Datasets:

CL-From-Nothing
/

FrozenLake-Hard-Trajectories

license: mit
task_categories:
  - reinforcement-learning
tags:
  - frozen-lake
  - 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

FrozenLake-Hard-Trajectories

Rollout trajectories for a hard variant of FrozenLake generated using the RAGEN framework.

Models

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

Environment Settings

Setting	Value
Grid size	8×8 (vs 4×4 in standard FrozenLake)
Frozen tile probability `p`	0.8 (~13 holes per map)
Movement success rate	0.8 (80% intended, 20% random slip)
Coordinate range	(0, 0) to (7, 7) (zero-indexed)
Max turns per episode	10
Actions per turn	up to 2 (`max_actions_per_turn=2`)

Note: 1 turn = up to 2 actions. All turn-based statistics count LLM calls.

Data Scale

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

Cutoff

Teacher average turns = 3.08 → cutoff = 2 turns (≤4 actions)

Note: floor(3.08 / 2) = 1, but cutoff was set to 2 to give a more meaningful prefix length given the harder environment.

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	0.068	0.150	4.42
Step 4: student→teacher	0.158	0.330	3.78
Step 5: teacher→student	0.167	0.316	3.37
Teacher only	0.268	0.466	3.08

The hard environment compresses the gap between all settings — even the teacher only achieves 26.8% success, leaving limited room for the continuation methods to recover. Both Step 4 and Step 5 roughly double the student baseline (6.8% → ~16%).

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

from datasets import load_dataset

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