scripts/inspect_rollout.py

"""
Inspect a saved rollout pkl file from RAGEN evaluation.

Usage:
    python scripts/inspect_rollout.py --path results/eval_multi/.../val_rollouts_*.pkl
    python scripts/inspect_rollout.py --path ... --n 5          # show first 5 episodes
    python scripts/inspect_rollout.py --path ... --idx 3        # show episode 3
    python scripts/inspect_rollout.py --path ... --success_only # only show successful episodes
    python scripts/inspect_rollout.py --path ... --summary      # summary stats only
"""

import argparse
import pickle
import sys
import os

sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "../verl"))


def load_pkl(path):
    with open(path, "rb") as f:
        return pickle.load(f)


def get_episodes(data):
    nb = data.non_tensor_batch
    messages_list = nb["messages_list"]
    # turn_counts may not exist (e.g. stitch rollouts); derive from messages
    if "turn_counts" in nb:
        turn_counts = nb["turn_counts"]
    else:
        turn_counts = [sum(1 for m in msgs if m["role"] == "assistant") for msgs in messages_list]

    episodes = []
    for i, (msgs, n_turns) in enumerate(zip(messages_list, turn_counts)):
        # collect intermediate rewards from user messages
        rewards = []
        for msg in msgs:
            if msg["role"] == "user" and "Reward:" in msg["content"]:
                lines = msg["content"].strip().splitlines()
                for line in lines:
                    if line.startswith("Reward:"):
                        try:
                            rewards.append(float(line.replace("Reward:", "").strip()))
                        except ValueError:
                            pass
                        break

        # infer success from last state: look for "√" (player on goal) in last assistant or user msg
        success = False
        for msg in reversed(msgs):
            content = msg["content"]
            if "√" in content or "player_on_goal" in content.lower():
                success = True
                break
            # also check last user message for final reward=1
            if msg["role"] == "user" and "Reward:" in content:
                for line in content.splitlines():
                    if line.startswith("Reward:"):
                        try:
                            if float(line.replace("Reward:", "").strip()) > 0:
                                success = True
                        except ValueError:
                            pass
                break

        episodes.append({
            "idx": i,
            "messages": msgs,
            "n_turns": n_turns,
            "rewards": rewards,
            "success": success,
        })
    return episodes


def print_summary(data, episodes):
    n = len(episodes)
    n_success = sum(e["success"] for e in episodes)
    avg_turns = sum(e["n_turns"] for e in episodes) / n

    # prefer aggregate metrics from meta_info if available
    metrics = (data.meta_info or {}).get("metrics", {})
    env_tag = next((k.split("/")[0] for k in metrics if "/success" in k), None)

    print(f"{'='*60}")
    if env_tag and f"{env_tag}/success" in metrics:
        print(f"[from meta_info] {env_tag}")
        print(f"  success       : {metrics[f'{env_tag}/success']:.1%}")
        pass_val = next((metrics[f"{env_tag}/{k}"] for k in ("pass@1", "pass@5", "pass@10") if f"{env_tag}/{k}" in metrics), "N/A")
        pass_key = next((k for k in ("pass@1", "pass@5", "pass@10") if f"{env_tag}/{k}" in metrics), "pass@1")
        print(f"  {pass_key:<14}: {pass_val}")
        print(f"  num_actions   : {metrics.get(f'{env_tag}/num_actions', 'N/A'):.2f}")
        print(f"  action_valid  : {metrics.get(f'{env_tag}/action_is_valid', 'N/A'):.1%}")
        print(f"  response_len  : {metrics.get('response_length', 'N/A'):.1f}")
    else:
        print(f"Total episodes : {n}")
        print(f"Success (inferred): {n_success} / {n}  ({100*n_success/n:.1f}%)")
    print(f"Avg turns      : {avg_turns:.2f}")
    print(f"{'='*60}")


def print_episode(ep, max_content_len=500):
    print(f"\n{'='*60}")
    status = "SUCCESS" if ep["success"] else "FAILED"
    print(f"Episode {ep['idx']}  [{status}]  turns={ep['n_turns']}  rewards={ep['rewards']}")
    print(f"{'='*60}")
    for i, msg in enumerate(ep["messages"]):
        role = msg["role"].upper()
        content = msg["content"]
        if len(content) > max_content_len:
            content = content[:max_content_len] + f"\n... [truncated {len(msg['content'])-max_content_len} chars]"
        print(f"\n[{role}]")
        print(content)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--path", required=True, help="Path to val_rollouts_*.pkl")
    parser.add_argument("--n", type=int, default=3, help="Number of episodes to show (default: 3)")
    parser.add_argument("--idx", type=int, default=None, help="Show a specific episode by index")
    parser.add_argument("--success_only", action="store_true", help="Only show successful episodes")
    parser.add_argument("--fail_only", action="store_true", help="Only show failed episodes")
    parser.add_argument("--summary", action="store_true", help="Show summary stats only")
    parser.add_argument("--max_len", type=int, default=500, help="Max chars per message to display")
    args = parser.parse_args()

    data = load_pkl(args.path)
    episodes = get_episodes(data)

    print_summary(data, episodes)

    if args.summary:
        return

    if args.idx is not None:
        print_episode(episodes[args.idx], args.max_len)
        return

    pool = episodes
    if args.success_only:
        pool = [e for e in episodes if e["success"]]
        print(f"Showing {min(args.n, len(pool))} successful episodes")
    elif args.fail_only:
        pool = [e for e in episodes if not e["success"]]
        print(f"Showing {min(args.n, len(pool))} failed episodes")

    for ep in pool[:args.n]:
        print_episode(ep, args.max_len)


if __name__ == "__main__":
    main()