---
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - multi-agent-path-finding
  - mapf
  - planning
  - llm-benchmark
pretty_name: MAPF-FrozenLake Benchmark
size_categories:
  - 1K<n<10K
configs:
  - config_name: benchmark_wr025
    data_files:
      - split: 3_agents
        path: benchmark_wr025/3_agents.jsonl
      - split: 4_agents
        path: benchmark_wr025/4_agents.jsonl
      - split: 5_agents
        path: benchmark_wr025/5_agents.jsonl
  - config_name: benchmark_wr050
    data_files:
      - split: 3_agents
        path: benchmark_wr050/3_agents.jsonl
      - split: 4_agents
        path: benchmark_wr050/4_agents.jsonl
      - split: 5_agents
        path: benchmark_wr050/5_agents.jsonl
  - config_name: benchmark_wr075
    data_files:
      - split: 3_agents
        path: benchmark_wr075/3_agents.jsonl
      - split: 4_agents
        path: benchmark_wr075/4_agents.jsonl
      - split: 5_agents
        path: benchmark_wr075/5_agents.jsonl
---

# MAPF-FrozenLake Benchmark

Evaluation benchmark for the paper
[From Trainee to Trainer: LLM-Designed Training Environment for RL with Multi-Agent Reasoning](https://arxiv.org/abs/2606.17682)

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

```python
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](https://github.com/LARK-AI-Lab/Trainee-to-Trainer)
repo root &mdash; 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:

```bash
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:

```bash
# 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.