scripts/README_EVAL_SCRIPTS.md

Evaluation Scripts

Quick reference for running evaluations on Sokoban environment.

Quick Start

Evaluate Qwen 2.5 7B (Default)

# 128 trajectories, GPU 0
bash scripts/eval_qwen_7b_sokoban.sh

# 256 trajectories
bash scripts/eval_qwen_7b_sokoban.sh 256

# Custom GPU
bash scripts/eval_qwen_7b_sokoban.sh 128 1

Output: outputs/qwen-2.5-7b-sokoban-128.jsonl

Evaluate Any Qwen Model

# General usage
bash scripts/eval_qwen_sokoban.sh <version> [num_traj] [gpu]

# Examples
bash scripts/eval_qwen_sokoban.sh 3B 128 0
bash scripts/eval_qwen_sokoban.sh 7B 256 1
bash scripts/eval_qwen_sokoban.sh 14B 128 0,1  # Multi-GPU

Output: outputs/qwen-<version>-sokoban-<num>.jsonl

Batch Evaluation (All Models)

bash scripts/eval_batch.sh

Evaluates all Qwen models (0.5B, 1.5B, 3B, 7B, 14B) sequentially.

Output: Multiple files in outputs/

Scripts Overview

eval_qwen_7b_sokoban.sh

• Purpose: Quick eval for Qwen 2.5 7B Instruct
• Args: [num_trajectories] [gpu_id]
• Default: 128 trajectories on GPU 0
• Output: outputs/qwen-2.5-7b-sokoban-{N}.jsonl

eval_qwen_sokoban.sh

• Purpose: Flexible eval for any Qwen model
• Args: <model_version> [num_trajectories] [gpu_id]
• Versions: 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B
• Output: outputs/qwen-{version}-sokoban-{N}.jsonl

eval_batch.sh

• Purpose: Evaluate multiple models in sequence
• Edit: Modify MODELS array to customize model list
• Output: One JSONL per model in outputs/

Output Format

All scripts output OpenAI-compatible JSONL:

{
  "custom_id": "traj_0",
  "messages": [
    {"role": "user", "content": "Grid state..."},
    {"role": "assistant", "content": "<think>...</think><ans>up</ans>"},
    {"role": "user", "content": "New grid... (reward: 0.0)"}
  ],
  "metadata": {
    "env_id": 0,
    "success": true,
    "total_reward": 1.0,
    "num_turns": 5
  }
}

Common Configurations

Trajectories

• 128: Good for quick eval (8 groups × 16)
• 256: Standard eval (16 groups × 16)
• 512: Thorough eval (32 groups × 16)

The script auto-rounds to nearest multiple of 16.

GPU Settings

• Single GPU: 0 or 1
• Multi-GPU: 0,1 or 0,1,2,3

Model Sizes

Model	VRAM	Recommended GPU
0.5B	~2GB	Any
1.5B	~4GB	RTX 3090
3B	~8GB	RTX 3090
7B	~16GB	A100 40GB
14B	~32GB	A100 80GB

Custom Evaluation

For full control, use the Python command directly:

python -m ragen.llm_agent.agent_proxy \
    --config-name eval \
    model_path="Qwen/Qwen2.5-7B-Instruct" \
    system.CUDA_VISIBLE_DEVICES="0" \
    es_manager.val.env_groups=8 \
    es_manager.val.group_size=16 \
    output.dir="outputs" \
    output.filename="custom-name.jsonl" \
    output.format=jsonl \
    output.append_timestamp=false

Troubleshooting

Out of Memory:

# Reduce context length in eval config
python -m ragen.llm_agent.agent_proxy --config-name eval \
    actor_rollout_ref.rollout.max_model_len=2048 \
    actor_rollout_ref.rollout.response_length=128

Model not found:

• Ensure model is downloaded or accessible via HuggingFace
• Check path format: Qwen/Qwen2.5-{size}B-Instruct

Slow evaluation:

• Use fewer trajectories for testing: bash scripts/eval_qwen_7b_sokoban.sh 32
• Enable greedy decoding: add actor_rollout_ref.rollout.val_kwargs.temperature=0