File size: 3,113 Bytes
67ba414 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 | """
Analyze rollout pkl: check whether successful episodes are all solved in turn=1.
Usage:
python scripts/test.py --path results/sft_data/sokoban_14B_10K/val_rollouts_20260410_215230.pkl
"""
import argparse
import os
import sys
import pickle
from collections import Counter
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_success_flags(data):
"""
Use rm_scores from the verifier directly.
rm_scores shape: [n_episodes, seq_len] — reward is placed at the last
valid token, all other positions are 0. Success = final nonzero reward > 0.
Falls back to message-based heuristic if rm_scores is absent.
"""
import torch
if data.batch is not None and "rm_scores" in data.batch:
rm = data.batch["rm_scores"] # [N, T]
# last nonzero reward per episode
final_rewards = []
for i in range(rm.shape[0]):
nonzero = rm[i].nonzero(as_tuple=False)
if len(nonzero) > 0:
final_rewards.append(rm[i, nonzero[-1]].item())
else:
final_rewards.append(0.0)
return [r > 0 for r in final_rewards]
# fallback
return [_is_success_from_msgs(msgs) for msgs in data.non_tensor_batch["messages_list"]]
def _is_success_from_msgs(msgs):
for m in reversed(msgs):
if m["role"] == "user" and "Reward:" in m["content"]:
for line in m["content"].splitlines():
if line.startswith("Reward:"):
try:
return float(line.replace("Reward:", "").strip()) > 0
except ValueError:
pass
break
return False
def count_turns(msgs):
"""Number of assistant turns in the episode."""
return sum(1 for m in msgs if m["role"] == "assistant")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--path", required=True)
args = parser.parse_args()
data = load_pkl(args.path)
msgs_list = data.non_tensor_batch["messages_list"]
success_flags = get_success_flags(data)
success_turn_dist = Counter()
fail_turn_dist = Counter()
for msgs, suc in zip(msgs_list, success_flags):
n_turns = count_turns(msgs)
if suc:
success_turn_dist[n_turns] += 1
else:
fail_turn_dist[n_turns] += 1
total = len(msgs_list)
total_success = sum(success_turn_dist.values())
print(f"Total episodes : {total}")
print(f"Total success : {total_success} ({100*total_success/total:.1f}%)")
print()
print("Turn distribution of SUCCESSFUL episodes:")
for t in sorted(success_turn_dist):
print(f" turns={t}: {success_turn_dist[t]} ({100*success_turn_dist[t]/total_success:.1f}%)")
print()
print("Turn distribution of FAILED episodes:")
for t in sorted(fail_turn_dist):
print(f" turns={t}: {fail_turn_dist[t]}")
if __name__ == "__main__":
main()
|