scripts/README_convert_rl_to_sft.md

FrozenLake RL Trajectories → LLM SFT Converter

This utility converts numeric RL testing trajectories collected from the FrozenLake wrapper into language trajectories suitable for supervised fine-tuning (SFT) of LLM agents.

It reconstructs the text grid from the one-hot observations saved during RL testing and formats per-turn prompts and responses to mirror RAGEN’s ContextManager chat format.

Path Layout

• Input: runs/<experiment>/trajectories/step_XXXXXX/trajectories.jsonl
• Output: runs/<experiment>/sft/step_XXXXXX_sft.jsonl python scripts/convert_rl_to_sft_frozenlake.py runs/FrozenLake__ppo_frozenlake__1__1762841111 --step step_993280 --include_failed

Each output line is a JSON object:

• messages: chat array with roles system, user, assistant, user (reward)
• meta: { episode_return, episode_success, global_step }

Requirements

• Python 3.8+
• Optional: pyyaml if you want the script to read config/envs.yaml and config/base.yaml for env_instruction, action_sep, and enable_think. If not installed, built-in defaults are used.

Usage

Convert the latest step under the run directory:

python scripts/convert_rl_to_sft_frozenlake.py runs/FrozenLake__ppo_frozenlake__1__1762841111

Convert a specific step directory:

python scripts/convert_rl_to_sft_frozenlake.py runs/FrozenLake__ppo_frozenlake__1__1762841111 --step step_993280

Include failed episodes (by default only successful episodes are kept):

python scripts/convert_rl_to_sft_frozenlake.py runs/FrozenLake__ppo_frozenlake_nochangeenv__1__1763561679 --step step_1986560

What the Converter Does

• Reconstructs the 4×4 grid text from the FrozenLake numeric state:
  • One-hot over 4 cell types plus 2 normalized coordinates for player position
  • Displays P for player on frozen, X for player in hole, and √ for player on goal
• Builds a chat-style prompt per episode:
  • system: "You're a helpful assistant."
  • user: env_instruction followed by per-turn blocks with State, remaining actions, and format constraints
  • assistant: Tagged action outputs: <think></think><answer>Action</answer> (or <answer>Action</answer> if think disabled)
  • user: Reward after each assistant response
• Maps RL actions (0..3) to RAGEN’s {Left, Down, Right, Up}
• Reads global_step from the step’s metrics.json if present

Configuration Hooks

• config/envs.yaml → FrozenLake.env_instruction, FrozenLake.max_tokens
• config/base.yaml → agent_proxy.action_sep (default ||), agent_proxy.enable_think (default True)

If these files are not present or pyyaml is not installed, the converter uses safe defaults.

Output Example (truncated)

{
  "messages": [
    {"role": "system", "content": "You're a helpful assistant. "},
    {"role": "user", "content": "You are solving the FrozenLake puzzle...\nTurn 1:\nState:\n__PO\n____\nG___\n____\nYou have 4 actions left..."},
    {"role": "assistant", "content": "<think></think><answer>Left</answer>"},
    {"role": "user", "content": "Reward:\n0.0\n"},
    ...
  ],
  "meta": {"episode_return": 1.0, "episode_success": true, "global_step": 993280}
}

Notes

• The converter currently targets FrozenLake; extending to Bandit and Sokoban is straightforward by adapting the state decoder.
• The per-turn structure follows the training-time prompt generator so SFT will be consistent with RL rollouts.