Spaces:
Sleeping
Sleeping
File size: 4,008 Bytes
40de84e | 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 | """Strip `handoff_notes` from all assistant action JSONs in SFT trajectories.
Phase B (models.py) removed `handoff_notes` from `PipelineAction`. The SFT
dataset still contained it in 289 of 390 assistant messages. Because the
parent `Action` class from openenv-core sets `extra="forbid"`, every call
to `PipelineAction(**action_data)` made by the GRPO reward / inference /
integration test code raises `ValidationError: extra_forbidden` when the
LLM emits an action with `handoff_notes` — which it WILL after SFT,
because the training data taught it that pattern.
This script normalises the dataset: each assistant message's content is a
single JSON object; we json.loads it, pop the offending key, and json.dumps
it back. Behavioural content (action sequences, roles, services, reasons)
is untouched.
A backup is written to data/sft_trajectories.jsonl.pre_strip.bak so the
operation is reversible.
Run from the project root:
python scripts/strip_handoff_notes.py
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
INPUT_PATH = Path("data/sft_trajectories.jsonl")
BACKUP_PATH = Path("data/sft_trajectories.jsonl.pre_strip.bak")
def _strip_in_action_json(content: str) -> tuple[str, bool]:
"""Try to parse `content` as a JSON action object. If it has handoff_notes,
drop it and return the re-serialised object. Otherwise return content
unchanged.
Returns (new_content, changed_bool).
"""
try:
obj = json.loads(content)
except json.JSONDecodeError:
return content, False
if not isinstance(obj, dict):
return content, False
if "handoff_notes" not in obj:
return content, False
del obj["handoff_notes"]
# Re-serialise compactly (matches existing single-line style); preserve
# non-ascii (e.g. em-dashes in `reason` fields) verbatim.
return json.dumps(obj, ensure_ascii=False), True
def main() -> int:
if not INPUT_PATH.exists():
print(f"ERROR: {INPUT_PATH} not found — run from project root", file=sys.stderr)
return 2
raw_lines = INPUT_PATH.read_text(encoding="utf-8").splitlines(keepends=True)
BACKUP_PATH.write_text("".join(raw_lines), encoding="utf-8")
print(f"Backup written: {BACKUP_PATH}")
output_lines: list[str] = []
total_records = 0
total_asst_msgs = 0
modified_asst_msgs = 0
total_changes = 0
bad_record_lines = 0
for line in raw_lines:
stripped = line.strip()
if not stripped or stripped.startswith("#"):
output_lines.append(line)
continue
try:
rec = json.loads(stripped)
except json.JSONDecodeError as e:
print(f"WARN: bad JSON line, preserving as-is: {e}")
output_lines.append(line)
bad_record_lines += 1
continue
total_records += 1
for msg in rec.get("messages", []):
if msg.get("role") != "assistant":
continue
total_asst_msgs += 1
new_content, changed = _strip_in_action_json(msg.get("content", ""))
if changed:
msg["content"] = new_content
modified_asst_msgs += 1
total_changes += 1
# Preserve trailing newline shape: existing file uses LF after each record.
output_lines.append(json.dumps(rec, ensure_ascii=False) + "\n")
INPUT_PATH.write_text("".join(output_lines), encoding="utf-8")
print(f"Total records processed: {total_records}")
print(f"Total assistant messages: {total_asst_msgs}")
print(f"Assistant messages modified: {modified_asst_msgs}")
print(f"Total handoff_notes occurrences removed: {total_changes}")
if bad_record_lines:
print(f"WARN: {bad_record_lines} non-JSON non-comment lines preserved verbatim")
return 0
if __name__ == "__main__":
sys.exit(main())
|