File size: 19,647 Bytes
7fde66e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 | """Convert public agentic trajectory datasets to OpenAI-messages SFT samples
matching the pi harness prompt surface (system prompt + read/bash/edit/write tools).
Sources:
1. MaxDevv/real-pi-coding-agent-traces-sessions (native pi tool traces)
2. R2E-Gym/R2EGym-SFT-Trajectories (XML <function=...> format)
3. TIGER-Lab/SWE-Next-SFT-Trajectories (XML, has tool role)
4. SWE-Gym/OpenHands-SFT-Trajectories (XML format)
5. nvidia/Nemotron-Terminal-Corpus skill_based_* (Terminus-2 JSON format)
Output: HF dataset with columns {messages, tools, source} saved to disk.
Only raw public hub datasets are used. Length-filtered with the Qwen3.5 tokenizer.
"""
import argparse
import glob
import json
import os
import re
import sys
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from pi_prompt import TOOLS, make_system_prompt
from datasets import Dataset, load_dataset
CORE_TOOLS = {"read", "bash", "edit", "write"}
MAX_TOOL_RESPONSE_CHARS = 30000
# ---------------------------------------------------------------------------
# helpers
# ---------------------------------------------------------------------------
def _trunc(s: str, n: int = MAX_TOOL_RESPONSE_CHARS) -> str:
if len(s) <= n:
return s
keep = n // 2
return s[:keep] + f"\n[... {len(s) - n} chars truncated ...]\n" + s[-keep:]
def _sample(messages, source, cwd="/testbed", date="2026-08-15"):
return {
"messages": [{"role": "system", "content": make_system_prompt(date=date, cwd=cwd)}] + messages,
"tools": json.dumps(TOOLS),
"source": source,
}
# ---------------------------------------------------------------------------
# 1. pi traces
# ---------------------------------------------------------------------------
def convert_pi_traces(hf_home: str, max_sessions: int | None = None):
snap = glob.glob(
os.path.join(hf_home, "hub", "datasets--MaxDevv--real-pi-coding-agent-traces-sessions", "snapshots", "*")
)[0]
files = sorted(glob.glob(os.path.join(snap, "*.jsonl")))
if max_sessions:
files = files[:max_sessions]
samples = []
stats = {"sessions": 0, "episodes": 0, "drop_tool": 0, "drop_empty": 0,
"drop_nostop": 0, "drop_bashexec": 0, "drop_role": 0}
for path in files:
stats["sessions"] += 1
events = []
with open(path) as f:
for line in f:
try:
events.append(json.loads(line))
except json.JSONDecodeError:
continue
msgs = [e["message"] for e in events if e.get("type") == "message" and "message" in e]
# split into episodes at user-message boundaries
episodes = []
cur = []
for m in msgs:
if m.get("role") == "user" and cur:
episodes.append(cur)
cur = []
cur.append(m)
if cur:
episodes.append(cur)
for ep in episodes:
out = []
ok = True
has_tool = False
final_stop = False
for m in ep:
role = m.get("role")
if role == "user":
text = "".join(b.get("text", "") for b in m.get("content", []) if isinstance(b, dict))
if not text.strip():
ok = False
break
out.append({"role": "user", "content": text})
elif role == "assistant":
content_parts, think_parts, tool_calls = [], [], []
stop = m.get("stopReason")
for b in m.get("content", []):
if not isinstance(b, dict):
continue
bt = b.get("type")
if bt == "text":
content_parts.append(b.get("text", ""))
elif bt == "thinking":
think_parts.append(b.get("thinking", b.get("text", "")))
elif bt == "toolCall":
name = b.get("name")
if name not in CORE_TOOLS:
ok = False
break
tool_calls.append({
"id": b.get("id", f"call_{len(tool_calls)}"),
"type": "function",
"function": {"name": name, "arguments": json.dumps(b.get("arguments", {}))},
})
has_tool = True
if not ok:
break
msg = {"role": "assistant", "content": "".join(content_parts)}
if think_parts:
msg["reasoning_content"] = "".join(think_parts)
if tool_calls:
msg["tool_calls"] = tool_calls
if not msg["content"] and not tool_calls and not think_parts:
continue
out.append(msg)
if stop == "stop":
final_stop = True
elif role == "toolResult":
if m.get("toolName") not in CORE_TOOLS:
ok = False
break
text = "".join(b.get("text", "") for b in m.get("content", []) if isinstance(b, dict))
out.append({
"role": "tool",
"tool_call_id": m.get("toolCallId", "call_0"),
"content": _trunc(text),
})
elif role == "bashExecution":
ok = False
stats["drop_bashexec"] += 1
break
else:
stats["drop_role"] += 1
if not ok:
stats["drop_tool"] += 1
continue
if not has_tool:
stats["drop_empty"] += 1
continue
if not final_stop or not out or out[-1]["role"] != "assistant":
stats["drop_nostop"] += 1
continue
stats["episodes"] += 1
samples.append(_sample(out, "pi_traces", cwd="/app"))
return samples, stats
# ---------------------------------------------------------------------------
# 2/3/4. XML <function=...> traces (R2EGym, SWE-Next, OpenHands)
# ---------------------------------------------------------------------------
FUNC_RE = re.compile(r"<function=([a-zA-Z_]+)>\s*(.*?)\s*</function>", re.DOTALL)
PARAM_RE = re.compile(r"<parameter=([a-zA-Z_]+)>\s*(.*?)\s*</parameter>", re.DOTALL)
def parse_xml_call(text: str):
"""Return (pre_text, func_name, params) or None."""
m = FUNC_RE.search(text)
if not m:
return None
pre = text[: m.start()].strip()
fname = m.group(1)
params = {p.group(1): p.group(2) for p in PARAM_RE.finditer(m.group(2))}
return pre, fname, params
def map_xml_tool(fname: str, params: dict):
"""Map XML tool to (name, args) for pi tools, or 'finish', or None to drop episode."""
if fname in ("file_editor", "str_replace_editor"):
cmd = params.get("command", "")
path = params.get("path", "")
if cmd == "view":
return ("read", {"path": path})
if cmd == "create":
return ("write", {"path": path, "content": params.get("file_text", "")})
if cmd == "str_replace":
return ("edit", {"path": path, "edits": [{"oldText": params.get("old_str", ""), "newText": params.get("new_str", "")}]})
return None # insert / undo_edit -> drop episode
if fname == "execute_bash":
return ("bash", {"command": params.get("cmd", params.get("command", ""))})
if fname == "search":
term = params.get("search_term", "").replace("'", "'\\''")
path = params.get("path", ".")
return ("bash", {"command": f"grep -rn '{term}' {path} | head -50"})
if fname == "finish":
return ("finish", {})
return None
RESULT_PREFIX_RE = re.compile(
r"^(Execution output of \[[^\]]+\]|EXECUTION RESULT of \[[^\]]+\]|Exit code: \d+)\s*:?\s*\n?",
re.IGNORECASE,
)
def strip_result_prefix(text: str) -> str:
prev = None
while prev != text:
prev = text
text = RESULT_PREFIX_RE.sub("", text)
return text.strip()
def convert_xml_traces(dataset_name: str, split: str, source: str, max_samples: int | None = None,
config: str | None = None):
ds = load_dataset(dataset_name, config, split=split) if config else load_dataset(dataset_name, split=split)
samples = []
stats = {"in": len(ds), "kept": 0, "drop_parse": 0, "drop_tool": 0, "drop_flow": 0}
call_id = 0
for ex in ds:
raw = ex["messages"]
# strip leading system message (we inject our own)
if raw and raw[0]["role"] == "system":
raw = raw[1:]
out = []
ok = True
i = 0
pending_call = None # tool_call awaiting its result message
while i < len(raw):
m = raw[i]
role = m["role"]
if role == "user" or role == "tool":
content = m["content"]
if pending_call is not None:
# this is a tool result
text = strip_result_prefix(content)
out.append({"role": "tool", "tool_call_id": pending_call, "content": _trunc(text)})
pending_call = None
else:
out.append({"role": "user", "content": content})
i += 1
elif role == "assistant":
parsed = parse_xml_call(m["content"])
if parsed is None:
# plain assistant text (e.g. final message without call)
if m["content"].strip():
out.append({"role": "assistant", "content": m["content"].strip()})
i += 1
continue
pre, fname, params = parsed
mapped = map_xml_tool(fname, params)
if mapped is None:
ok = False
stats["drop_tool"] += 1
break
name, args = mapped
if name == "finish":
text = pre or "The task is complete."
out.append({"role": "assistant", "content": text})
i += 1
continue
call_id += 1
cid = f"call_{call_id}"
msg = {"role": "assistant", "content": pre,
"tool_calls": [{"id": cid, "type": "function",
"function": {"name": name, "arguments": json.dumps(args)}}]}
out.append(msg)
pending_call = cid
i += 1
else:
i += 1
if not ok:
continue
# validate flow: alternating, no dangling call, ends with assistant
if pending_call is not None or not out or out[-1]["role"] != "assistant":
stats["drop_flow"] += 1
continue
if not any("tool_calls" in mm for mm in out):
stats["drop_flow"] += 1
continue
stats["kept"] += 1
samples.append(_sample(out, source, cwd="/testbed"))
if max_samples and len(samples) >= max_samples:
break
return samples, stats
# ---------------------------------------------------------------------------
# 5. Terminus-2 JSON traces (Nemotron Terminal Corpus skill_based_*)
# ---------------------------------------------------------------------------
def convert_terminus(config: str, source: str, max_samples: int, seed: int = 0):
ds = load_dataset("nvidia/Nemotron-Terminal-Corpus", config, split="train")
if max_samples and len(ds) > max_samples:
ds = ds.shuffle(seed=seed).select(range(max_samples))
samples = []
stats = {"in": len(ds), "kept": 0, "drop_task": 0, "drop_parse": 0, "drop_flow": 0}
call_id = 0
for ex in ds:
conv = ex["conversations"]
if not conv or conv[0]["role"] != "user":
stats["drop_flow"] += 1
continue
first = conv[0]["content"]
tm = re.search(r"\n\nTask Description:\n", first)
if not tm:
stats["drop_task"] += 1
continue
task = first[tm.end():]
# strip trailing terminal-state scaffolding
task = re.split(r"\nCurrent terminal state:|\nCurrent Terminal Screen:", task)[0].strip()
if not task:
stats["drop_task"] += 1
continue
out = [{"role": "user", "content": task}]
ok = True
i = 1
pending_call = None
while i < len(conv):
m = conv[i]
if m["role"] == "assistant":
content = m["content"]
think = ""
tm2 = re.match(r"\s*<think>(.*?)</think>\s*", content, re.DOTALL)
rest = content
if tm2:
think = tm2.group(1).strip()
rest = content[tm2.end():]
# parse JSON block
jstart = rest.find("{")
if jstart == -1:
# plain text assistant message
if rest.strip():
out.append({"role": "assistant", "content": rest.strip()})
i += 1
continue
try:
obj, _ = json.JSONDecoder().raw_decode(rest[jstart:])
except json.JSONDecodeError:
ok = False
stats["drop_parse"] += 1
break
analysis = (obj.get("analysis") or "").strip()
plan = (obj.get("plan") or "").strip()
commands = obj.get("commands") or []
task_complete = obj.get("task_complete", False)
text = "\n\n".join(p for p in (analysis, f"Plan: {plan}" if plan else "") if p)
msg = {"role": "assistant", "content": text}
if think:
msg["reasoning_content"] = think
if commands:
script = "".join(c.get("keystrokes", "") for c in commands if isinstance(c, dict))
if script.strip():
call_id += 1
cid = f"call_{call_id}"
msg["tool_calls"] = [{"id": cid, "type": "function",
"function": {"name": "bash", "arguments": json.dumps({"command": script})}}]
pending_call = cid
out.append(msg)
i += 1
if task_complete:
break
elif m["role"] == "user":
content = m["content"]
content = re.sub(r"^New Terminal Output:\n?", "", content)
if pending_call is not None:
out.append({"role": "tool", "tool_call_id": pending_call, "content": _trunc(content)})
pending_call = None
else:
out.append({"role": "user", "content": content})
i += 1
else:
i += 1
if not ok:
continue
if pending_call is not None or not out or out[-1]["role"] != "assistant":
stats["drop_flow"] += 1
continue
if not any("tool_calls" in mm for mm in out):
stats["drop_flow"] += 1
continue
stats["kept"] += 1
samples.append(_sample(out, source, cwd="/app"))
return samples, stats
# ---------------------------------------------------------------------------
# length filter + main
# ---------------------------------------------------------------------------
def _args_to_dict(messages):
out = []
for m in messages:
m = dict(m)
if m.get("tool_calls"):
tcs = []
for tc in m["tool_calls"]:
tc = dict(tc)
fn = dict(tc["function"])
if isinstance(fn["arguments"], str):
fn["arguments"] = json.loads(fn["arguments"])
tc["function"] = fn
tcs.append(tc)
m["tool_calls"] = tcs
out.append(m)
return out
def token_len_filter(samples, tokenizer, max_tokens: int):
kept, dropped = [], 0
for s in samples:
try:
r = tokenizer.apply_chat_template(
_args_to_dict(s["messages"]), tools=json.loads(s["tools"]), add_generation_prompt=False
)
n = len(r["input_ids"])
except Exception:
dropped += 1
continue
if n <= max_tokens:
s["n_tokens"] = n
kept.append(s)
else:
dropped += 1
return kept, dropped
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--out", required=True)
ap.add_argument("--max-tokens", type=int, default=15000)
ap.add_argument("--max-pi-sessions", type=int, default=None)
ap.add_argument("--max-terminus", type=int, default=8000)
ap.add_argument("--max-r2e", type=int, default=None)
ap.add_argument("--max-swenext", type=int, default=None)
ap.add_argument("--skip-terminus", action="store_true")
ap.add_argument("--skip-pi", action="store_true")
ap.add_argument("--skip-xml", action="store_true")
args = ap.parse_args()
hf_home = os.environ["HF_HOME"]
all_samples = []
if not args.skip_pi:
s, st = convert_pi_traces(hf_home, args.max_pi_sessions)
print(f"[pi_traces] {st}", flush=True)
all_samples += s
if not args.skip_xml:
s, st = convert_xml_traces("R2E-Gym/R2EGym-SFT-Trajectories", "train", "r2egym", args.max_r2e)
print(f"[r2egym] {st}", flush=True)
all_samples += s
s, st = convert_xml_traces("TIGER-Lab/SWE-Next-SFT-Trajectories", "train", "swenext", args.max_swenext)
print(f"[swenext] {st}", flush=True)
all_samples += s
s, st = convert_xml_traces("SWE-Gym/OpenHands-SFT-Trajectories", "train.success.oss", "openhands")
print(f"[openhands] {st}", flush=True)
all_samples += s
if not args.skip_terminus:
for cfg in ("skill_based_easy", "skill_based_medium", "skill_based_mixed"):
s, st = convert_terminus(cfg, f"terminus_{cfg}", args.max_terminus // 3)
print(f"[{cfg}] {st}", flush=True)
all_samples += s
print(f"total before length filter: {len(all_samples)}", flush=True)
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.5-9B-Base")
all_samples, dropped = token_len_filter(all_samples, tok, args.max_tokens)
print(f"length filter: kept {len(all_samples)}, dropped {dropped}", flush=True)
import random
random.seed(0)
random.shuffle(all_samples)
ds = Dataset.from_list(all_samples)
ds.save_to_disk(args.out)
print(f"saved {len(ds)} samples to {args.out}", flush=True)
from collections import Counter
print(Counter(ds["source"]), flush=True)
lens = ds["n_tokens"]
import numpy as np
print(f"tokens: mean {np.mean(lens):.0f} p50 {np.percentile(lens,50):.0f} p90 {np.percentile(lens,90):.0f} max {max(lens)}", flush=True)
if __name__ == "__main__":
main()
|