File size: 7,768 Bytes
09a4637 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 | """
Parse raw Search-agent rollout trajectories (txt format) into structured JSON.
Search-agent rollouts use a different on-disk format than ALFWorld/WebShop:
- The trajectory is encoded as a History block followed by 'Now it's your
turn' and a final action.
- Each step is `Step <i>: <search/answer tag>` plus an optional
`<information>...</information>` block (the retrieved document).
- 'Reward' marks the final reward (1.0 for success, else failure).
Output schema (one record per .txt file):
{
"env_id": "<envFolder>_<filename>",
"task": "<question>",
"data_source": "nq | popqa | 2wikimultihopqa | triviaqa | hotpotqa | musique | bamboogle",
"type": "all_success" | "all_fail",
"outcome": "Success" | "Failure",
"reward": float,
"trajectories": [[step_dict, ...]]
}
Each step contains `step_id`, `action` (the <search>/<answer> tag), and
`memory_context` — a string in the same format the agent would see at
inference, replaying prior steps' actions and information snippets.
Filtering:
- Per data_source, keep ALL successes and only the first
`failure_ratio * n_success` failures (sorted by env_id), to roughly
balance the dataset.
Usage:
python parse_search.py \\
--input_dir /path/to/trajectories \\
--output_file processed_trajectories_search.json \\
--failure_ratio 0.5
"""
import argparse
import glob
import json
import os
import re
PATTERN_QUESTION = re.compile(r"Your question:\s*(.*?)(?:\n|$)", re.DOTALL)
PATTERN_REWARD = re.compile(r"Rewards?:\s*([\d\.]+)")
PATTERN_HISTORY_STEP = re.compile(
r"Step (\d+):\s*"
r"(<(?:search|answer)>.*?</(?:search|answer)>)"
r"\s*"
r"(?:<information>(.*?)</information>)?",
re.DOTALL,
)
PATTERN_FINAL_ACTION = re.compile(
r"Action \d+:\s*(<(?:search|answer)>.*?</(?:search|answer)>)", re.DOTALL
)
def infer_data_source(file_path: str) -> str:
"""Infer the QA dataset from the file index. Adjust the ranges to match
your rollout numbering scheme."""
base_name = os.path.basename(file_path)
try:
idx = int(os.path.splitext(base_name)[0])
except ValueError:
return "unknown"
ranges = [
(0, 80, "nq"),
(81, 160, "popqa"),
(161, 240, "2wikimultihopqa"),
(241, 320, "triviaqa"),
(321, 400, "hotpotqa"),
(401, 480, "musique"),
]
for lo, hi, src in ranges:
if lo <= idx <= hi:
return src
return "bamboogle"
def parse_search_trajectory_file(file_path: str):
try:
with open(file_path, "r", encoding="utf-8") as f:
content = f.read()
except Exception as e:
print(f"[Error] reading {file_path}: {e}")
return None
q_match = PATTERN_QUESTION.search(content)
task_desc = q_match.group(1).strip() if q_match else "Unknown Task"
r_match = PATTERN_REWARD.search(content)
reward = float(r_match.group(1)) if r_match else 0.0
is_success = reward == 1.0
steps = []
accumulated_history_str = ""
history_matches = list(
re.finditer(r"History:\s*\n(.*?)\nNow it's your turn", content, re.DOTALL)
)
if history_matches:
last_history_block = history_matches[-1].group(1)
for m in PATTERN_HISTORY_STEP.finditer(last_history_block):
step_id = int(m.group(1))
action_str = m.group(2).strip()
raw_obs_str = m.group(3).strip() if m.group(3) else ""
steps.append(
{
"step_id": step_id,
"action": action_str,
"memory_context": accumulated_history_str,
"reward": 0.0,
"done": False,
}
)
obs_tag = f" <information>{raw_obs_str}</information>" if raw_obs_str else ""
accumulated_history_str += f"Step {step_id}:{action_str}{obs_tag}\n"
last_turn_index = content.rfind("Now it's your turn")
if last_turn_index != -1:
last_section = content[last_turn_index:]
action_match = PATTERN_FINAL_ACTION.search(last_section)
if action_match:
final_action = action_match.group(1).strip()
if not steps or steps[-1]["action"] != final_action:
steps.append(
{
"step_id": len(steps),
"action": final_action,
"memory_context": accumulated_history_str,
"reward": reward,
"done": True,
}
)
if not steps:
single_action = PATTERN_FINAL_ACTION.search(content)
if single_action:
steps.append(
{
"step_id": 0,
"action": single_action.group(1).strip(),
"memory_context": "",
"reward": reward,
"done": True,
}
)
else:
return None
env_folder = os.path.basename(os.path.dirname(file_path))
file_name = os.path.basename(file_path)
return {
"env_id": f"{env_folder}_{file_name}",
"task": task_desc,
"data_source": infer_data_source(file_path),
"type": "all_success" if is_success else "all_fail",
"outcome": "Success" if is_success else "Failure",
"reward": reward,
"trajectories": [steps],
}
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--input_dir", required=True, help="Root containing **/*.txt rollouts")
parser.add_argument("--output_file", required=True)
parser.add_argument(
"--failure_ratio",
type=float,
default=0.5,
help="Per data_source: keep failures up to this fraction of successes.",
)
args = parser.parse_args()
files = glob.glob(os.path.join(args.input_dir, "**", "*.txt"), recursive=True)
print(f"Found {len(files)} files...")
successes_by_source, failures_by_source = {}, {}
stats = {}
for i, fpath in enumerate(files):
parsed = parse_search_trajectory_file(fpath)
if parsed is None:
continue
source = parsed.get("data_source", "unknown")
stats.setdefault(source, {"Success": 0, "Failure": 0})
stats[source][parsed["outcome"]] += 1
target = successes_by_source if parsed["outcome"] == "Success" else failures_by_source
target.setdefault(source, []).append(parsed)
if (i + 1) % 100 == 0:
print(f"Processed {i+1} files...")
final_dataset = []
for source in sorted(stats):
successes = successes_by_source.get(source, [])
failures = sorted(failures_by_source.get(source, []), key=lambda x: x["env_id"])
keep_failures = int(len(successes) * args.failure_ratio)
final_dataset.extend(successes)
final_dataset.extend(failures[:keep_failures])
final_dataset.sort(key=lambda x: x["env_id"])
n_success = sum(1 for e in final_dataset if e["outcome"] == "Success")
n_fail = sum(1 for e in final_dataset if e["outcome"] == "Failure")
os.makedirs(os.path.dirname(os.path.abspath(args.output_file)) or ".", exist_ok=True)
with open(args.output_file, "w", encoding="utf-8") as f:
json.dump(final_dataset, f, indent=2, ensure_ascii=False)
print(f"Saved {len(final_dataset)} entries (Success={n_success}, Fail={n_fail}).")
if stats:
print("Per data_source:")
for source in sorted(stats):
s = len(successes_by_source.get(source, []))
kept = int(s * args.failure_ratio)
print(f" {source}: Success {s}, Fail kept {kept}")
if __name__ == "__main__":
main()
|