Initial upload: MAPF-FrozenLake eval benchmark

README.md

@@ -1,3 +1,111 @@
 ---
 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: wr025
+    data_files:
+      - split: 3_agents
+        path: wr025/3_agents.jsonl
+      - split: 4_agents
+        path: wr025/4_agents.jsonl
+      - split: 5_agents
+        path: wr025/5_agents.jsonl
+  - config_name: wr050
+    data_files:
+      - split: 3_agents
+        path: wr050/3_agents.jsonl
+      - split: 4_agents
+        path: wr050/4_agents.jsonl
+      - split: 5_agents
+        path: wr050/5_agents.jsonl
+  - config_name: wr075
+    data_files:
+      - split: 3_agents
+        path: wr075/3_agents.jsonl
+      - split: 4_agents
+        path: wr075/4_agents.jsonl
+      - split: 5_agents
+        path: 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://github.com/LARK-AI-Lab/Trainee-to-Trainer)**.
+
+Three configs (`wr025` / `wr050` / `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="wr075", split="5_agents")
+print(ds[0]["text"][:400])
+```
+
+## Run evaluation
+
+Place the benchmark under the
+[Trainee-to-Trainer](https://github.com/LARK-AI-Lab/Trainee-to-Trainer)
+repo root at one of `benchmark_wr025/` / `benchmark_wr050/` /
+`benchmark_wr075/`, with this exact layout:
+
+```
+benchmark_wr075/
+├── 3_agents/dataset_nl.jsonl
+├── 4_agents/dataset_nl.jsonl
+└── 5_agents/dataset_nl.jsonl
+```
+
+You can download + lay out everything in one go:
+
+```bash
+hf download LARK-Lab/MAPF-FrozenLake-Benchmark \
+  --repo-type dataset --local-dir /tmp/mapf_bench
+
+for wr in wr025 wr050 wr075; do
+  for n in 3 4 5; do
+    mkdir -p benchmark_${wr}/${n}_agents
+    cp /tmp/mapf_bench/${wr}/${n}_agents.jsonl \
+       benchmark_${wr}/${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.