scripts/convert_rl_to_sft_2048.py

#!/usr/bin/env python3
"""
Convert RL eval trajectories from Game 2048 into LLM SFT-ready chat data.
Matches the specific text format of the RAGEN/Maniskill environment runner.
"""

import argparse
import json
import math
from pathlib import Path
from typing import List, Tuple
import numpy as np
try:
    import yaml  # type: ignore
except Exception:
    yaml = None


# 2048 action lookup from config (0..3)
ACTION_LOOKUP_2048 = {0: "Up", 1: "Right", 2: "Down", 3: "Left"}


def load_env_instruction_and_cfg(repo_root: Path) -> Tuple[str, int, str, bool, int]:
    """Load 2048 env instruction and base agent_proxy configs.
    Returns: (instruction, max_tokens, action_sep, enable_think, max_actions)
    """
    # Defaults
    max_tokens = 8192  # Adjusted to match your log (was 64)
    action_sep = " || " # Adjusted spacing to match your log
    enable_think = True
    max_actions = 1000  # Default limit

    # We construct the instruction to strictly match the environment text
    # Note: The dynamic parts (like max_actions) are inserted here.
    instruction_template = (
        "You are playing the 2048 game on a 4x4 grid. Merge equal tiles by sliding Up, Right, Down, or Left.\n"
        "If a move is invalid (no tiles move), a small penalty is applied. Respond with a single action.\n"
        "Example: <answer>Up</answer>\n\n"
        "Your available actions are:\n"
        "Up, Right, Down, Left\n"
        "You can make up to {max_actions} actions, separated by the action separator \"{action_sep}\""
    )

    if yaml is not None:
        # envs.yaml
        envs_yaml = repo_root / "config" / "envs.yaml"
        if envs_yaml.exists():
            try:
                with open(envs_yaml, "r", encoding="utf-8") as f:
                    envs = yaml.safe_load(f)
                if isinstance(envs, dict) and "custom_envs" in envs and "game_2048" in envs["custom_envs"]:
                    e = envs["custom_envs"]["game_2048"]
                    # If you want to use the yaml instruction, uncomment below. 
                    # But for strict matching of your provided text, we prefer the hardcoded template above.
                    # instruction = e.get("env_instruction", instruction)
                    
                    # We still load configs
                    # max_tokens = int(e.get("max_tokens", max_tokens)) 
                    max_actions = int(e.get("max_actions_per_traj", max_actions))
            except Exception:
                pass

        # base.yaml
        base_yaml = repo_root / "config" / "base.yaml"
        if base_yaml.exists():
            try:
                with open(base_yaml, "r", encoding="utf-8") as f:
                    base_cfg = yaml.safe_load(f)
                ap = base_cfg.get("agent_proxy", {}) if isinstance(base_cfg, dict) else {}
                # action_sep = ap.get("action_sep", action_sep)
                enable_think = bool(ap.get("enable_think", enable_think))
            except Exception:
                pass

    instruction = instruction_template.format(max_actions=max_actions, action_sep=action_sep)
    return instruction, max_tokens, action_sep, enable_think, max_actions


def decode_state_to_grid_text(state: List[List[List[float]]]) -> str:
    """Decode 2048 CNN one-hot channels (C x 4 x 4) back to a grid string."""
    if not state or not isinstance(state, list):
        return ""
    # Expect shape (C, 4, 4)
    C = len(state)
    H = len(state[0]) if C > 0 else 0
    W = len(state[0][0]) if (C > 0 and H > 0) else 0
    grid_vals: List[List[int]] = [[0 for _ in range(W)] for __ in range(H)]
    for i in range(H):
        for j in range(W):
            max_c = 0
            max_v = -float("inf")
            for c in range(C):
                v = state[c][i][j]
                if v > max_v:
                    max_v = v
                    max_c = c
            if max_c <= 0:
                grid_vals[i][j] = 0
            else:
                try:
                    grid_vals[i][j] = int(2 ** max_c)
                except Exception:
                    grid_vals[i][j] = 0
    
    lines = ["Current 2048 Grid:"]
    for r in range(H):
        row_str = ", ".join(str(grid_vals[r][c]) for c in range(W))
        lines.append(f"Row {r+1}: [{row_str}]")
    
    return "\n".join(lines)


def build_messages_for_episode(
    states: List[List[List[List[float]]]],
    actions: List[int],
    rewards: List[float],
    instruction: str,
    max_tokens: int,
    action_sep: str,
    enable_think: bool,
    max_actions: int,
) -> List[dict]:
    messages = [
        {"role": "system", "content": "You're a helpful assistant. "},
        {"role": "user", "content": instruction},
    ]

    total_actions = len(actions)
    current_score = 0.0

    for t, state in enumerate(states):
        grid_text = decode_state_to_grid_text(state)
        actions_left = max(0, max_actions - t)
        
        # Matches: "<think> ... </think> <answer> ... </answer> with no extra text."
        # Note the space before 'with'.
        format_prompt = (
            "<think> [Your thoughts] </think> <answer> [your answer] </answer>"
            if enable_think
            else "<answer> [your answer] </answer>"
        )
        
        # Matches: "Max response length: 4096 words (tokens)."
        length_prompt = f"Max response length: {max_tokens} words (tokens)."

        # Construct the User content block strictly matching the target format
        # Note: Your target text has blank lines represented by unicode non-breaking spaces or just empty lines.
        # We use standard \n for separation.
        
        turn_content = (
            f"\n\nTurn {t + 1}:\n"
            f"State:\n"
            f"{grid_text}\n\n"
            f"Valid Actions: 0(Up), 1(Right), 2(Down), 3(Left).\n"
            f"Goal: Merge same numbers to reach 2048.\n"
            f"Current Score: {int(current_score)}\n"
            f"What is your next move?\n"
            f"You have {actions_left} actions left. Always output: {format_prompt} "
            f"with no extra text. Strictly follow this format. {length_prompt}"
        )

        messages[-1]["content"] += turn_content

        if t < total_actions:
            a = actions[t]
            a_name = ACTION_LOOKUP_2048.get(int(a), str(a))
            if enable_think:
                assistant_text = f"<think> </think><answer>{a_name}</answer>"
            else:
                assistant_text = f"<answer>{a_name}</answer>"
            
            messages.append({"role": "assistant", "content": assistant_text})
            
            r = rewards[t] if t < len(rewards) else 0.0
            # Update score for the NEXT turn display
            current_score += r 
            
            messages.append({"role": "user", "content": f"Reward:\n{float(np.log2(r + 1.0)) * 0.1}"})

    # The last element is a user reward message for the final step; trim if needed for SFT
    return messages[:-1]


def find_latest_step_dir(traj_root: Path) -> Path:
    step_dirs = [p for p in traj_root.iterdir() if p.is_dir() and p.name.startswith("step_")]
    if not step_dirs:
        raise FileNotFoundError(f"No step_* directories under {traj_root}")
    step_dirs.sort(key=lambda p: int(p.name.split("_")[-1]))
    return step_dirs[-1]


def convert_file(step_dir: Path, output_dir: Path, repo_root: Path, include_failed: bool, max_actions_cap: int | None) -> Path:
    traj_path = step_dir / "trajectories.jsonl"
    metrics_path = step_dir / "metrics.json"
    if not traj_path.exists():
        raise FileNotFoundError(f"Missing trajectories.jsonl at {traj_path}")

    instruction, max_tokens, action_sep, enable_think, default_max_actions = load_env_instruction_and_cfg(repo_root)
    max_actions = int(max_actions_cap) if max_actions_cap is not None else int(default_max_actions)
    
    output_dir.mkdir(parents=True, exist_ok=True)
    out_path = output_dir / f"{step_dir.name}_sft.jsonl"

    # Read global step from metrics if available
    global_step = None
    if metrics_path.exists():
        try:
            with open(metrics_path, "r", encoding="utf-8") as f:
                m = json.load(f)
            global_step = m.get("global_step")
        except Exception:
            pass

    written = 0

    with open(traj_path, "r", encoding="utf-8") as fin, open(out_path, "w", encoding="utf-8") as fout:
        for line in fin:
            line = line.strip()
            if not line:
                continue
            traj = json.loads(line)

            ep_success = bool(traj.get("episode_success", False))
            if (not include_failed) and (not ep_success):
                continue

            states = traj.get("states", [])
            actions = traj.get("actions", [])
            rewards = traj.get("rewards", [])
            
            # Simple length check
            if len(actions) > max_actions:
                continue

            messages = build_messages_for_episode(
                states=states,
                actions=actions,
                rewards=rewards,
                instruction=instruction,
                max_tokens=max_tokens,
                action_sep=action_sep,
                enable_think=enable_think,
                max_actions=max_actions,
            )

            record = {
                "messages": messages,
                "meta": {
                    "episode_return": traj.get("episode_return", None),
                    "episode_success": ep_success,
                    "global_step": global_step,
                },
            }
            
            # Filter condition from your original script
            if traj.get("episode_return", 0) > 7000:
                fout.write(json.dumps(record, ensure_ascii=False) + "\n")
                written += 1
    import pdb;pdb.set_trace()
    if written == 0:
        # write empty file to indicate execution
        with open(out_path, "w", encoding="utf-8"):
            pass
    return out_path


def main():
    parser = argparse.ArgumentParser(description="Convert 2048 RL trajectories to LLM SFT chat JSONL")
    parser.add_argument("run_dir", help="Path to the run directory (contains trajectories/)")
    parser.add_argument("--step", default=None, help="Specific step directory name (e.g., step_993280)")
    parser.add_argument("--include_failed", action="store_true", help="Include failed episodes in SFT data")
    parser.add_argument("--max_actions", type=int, default=1000, help="Override max actions cap (default from envs.yaml)")
    args = parser.parse_args()

    repo_root = Path(__file__).resolve().parents[1]
    run_dir = Path(args.run_dir)
    traj_root = run_dir / "trajectories"
    if not traj_root.exists():
        raise FileNotFoundError(f"Not found trajectories directory: {traj_root}")

    step_dir = traj_root / args.step if args.step else find_latest_step_dir(traj_root)
    output_dir = run_dir / "sft"
    out_path = convert_file(step_dir=step_dir, output_dir=output_dir, repo_root=repo_root, include_failed=args.include_failed, max_actions_cap=args.max_actions)
    print(f"SFT data written to: {out_path}")


if __name__ == "__main__":
    main()