LARK-Lab/MAPF-FrozenLake-Benchmark

MAPF-FrozenLake Benchmark

Evaluation benchmark for the paper From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning

Three configs (benchmark_wr025 / benchmark_wr050 / benchmark_wr075) correspond to the wait-ratio threshold of the underlying CBS-optimal solution (higher = more inter-agent coordination required). Each config has three splits by agent count.

Load

from datasets import load_dataset

ds = load_dataset("LARK-Lab/MAPF-FrozenLake-Benchmark",
                  name="benchmark_wr075", split="5_agents")
print(ds[0]["text"][:400])

Run evaluation

Drop the downloaded folders directly into the Trainee-to-Trainer repo root — the directory names already match what the evaluators expect (benchmark_wr025/ / benchmark_wr050/ / benchmark_wr075/). Each one must contain <N>_agents/dataset_nl.jsonl.

One-shot download + layout:

hf download LARK-Lab/MAPF-FrozenLake-Benchmark \
  --repo-type dataset --local-dir /tmp/mapf_bench

for wr in benchmark_wr025 benchmark_wr050 benchmark_wr075; do
  for n in 3 4 5; do
    mkdir -p ${wr}/${n}_agents
    cp /tmp/mapf_bench/${wr}/${n}_agents.jsonl \
       ${wr}/${n}_agents/dataset_nl.jsonl
  done
done

Then run the evaluators shipped with the code repo:

# HuggingFace-format model
DATA_ROOT=benchmark_wr075 sbatch test_model_hf.sh \
    /path/to/model "3,4,5" "3,4,5,6,7,8,9,10" my_tag

# FSDP RL checkpoint
DATA_ROOT=benchmark_wr075 sbatch test_model_rl.sh \
    /path/to/outputs/.../global_step_XXX

# OpenAI-compatible API model
bash test_model_api.sh <endpoint> <model-id> <api-key>

Each run prints per-(agent-count, map-size) optimal-rate and accuracy at the end of its log.

License

MIT.