feat: export conversation trajectories to ShareGPT JSONL for LoRA fine-tuning
Some checks failed
Tests / test (pull_request) Has been skipped
Tests / lint (pull_request) Failing after 16s

Implements AutoLoRA Step 3 of 7: a script that reads Timmy's session logs
and chat history, groups entries into conversation trajectories, and writes
ShareGPT-compatible JSONL suitable for Hermes 4 LoRA fine-tuning.

Sources (priority order):
  1. logs/session_*.jsonl — rich logs with tool calls
  2. data/chat.db         — SQLite chat history fallback

Usage:
  python scripts/export_trajectories.py [--output ~/timmy-training-data.jsonl]
  python scripts/export_trajectories.py --validate-only --output <file>
  python scripts/export_trajectories.py --min-examples 100

Fixes #1102

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Alexander Whitestone
2026-03-23 14:20:04 -04:00
parent 852fec3681
commit a2f8989c39
2 changed files with 664 additions and 0 deletions

View File

@@ -0,0 +1,358 @@
#!/usr/bin/env python3
"""Export Claude conversation trajectories to ShareGPT JSONL format for LoRA fine-tuning.
Reads from two sources (in priority order):
1. logs/session_*.jsonl — rich logs with tool calls (preferred)
2. data/chat.db — SQLite chat history (fallback)
Output is a ShareGPT-compatible JSONL file where each line is one conversation:
{"conversations": [
{"from": "human", "value": "..."},
{"from": "gpt", "value": "...", "tool_calls": [...]},
{"from": "tool", "value": "..."},
{"from": "gpt", "value": "..."}
]}
Epic: #1091 Project Bannerlord — AutoLoRA Sovereignty Loop (Step 3 of 7)
Refs: #1102
"""
from __future__ import annotations
import argparse
import json
import sqlite3
import sys
from datetime import datetime, timedelta
from pathlib import Path
from typing import Any
# ── Constants ────────────────────────────────────────────────────────────────
REPO_ROOT = Path(__file__).resolve().parents[1]
DEFAULT_LOGS_DIR = REPO_ROOT / "logs"
DEFAULT_DB_PATH = REPO_ROOT / "data" / "chat.db"
DEFAULT_OUTPUT = Path.home() / "timmy-training-data.jsonl"
# Time gap that signals a new conversation boundary
CONVERSATION_GAP_MINUTES = 30
# Role mappings → ShareGPT "from" values
ROLE_MAP = {
"user": "human",
"timmy": "gpt",
"agent": "gpt",
"assistant": "gpt",
"system": "system",
}
# ── Session log reader ───────────────────────────────────────────────────────
def _parse_ts(ts: str) -> datetime | None:
"""Parse an ISO timestamp string, returning None on failure."""
try:
return datetime.fromisoformat(ts)
except (ValueError, TypeError):
return None
def _group_into_conversations(
entries: list[dict],
gap_minutes: int = CONVERSATION_GAP_MINUTES,
) -> list[list[dict]]:
"""Split a flat list of session entries into conversation windows.
A new conversation starts whenever there is a gap ≥ *gap_minutes* between
consecutive entries, or when the type sequence restarts with a user message
after an agent reply.
"""
if not entries:
return []
conversations: list[list[dict]] = []
current: list[dict] = []
last_ts: datetime | None = None
for entry in entries:
ts = _parse_ts(entry.get("timestamp", ""))
if last_ts is not None and ts is not None:
gap = ts - last_ts
if gap >= timedelta(minutes=gap_minutes):
if current:
conversations.append(current)
current = []
current.append(entry)
if ts is not None:
last_ts = ts
if current:
conversations.append(current)
return conversations
def _conversation_to_sharegpt(entries: list[dict]) -> dict[str, Any] | None:
"""Convert a list of session entries into a ShareGPT conversation dict.
Returns None if the conversation has fewer than 2 turns (not useful for
training).
"""
turns: list[dict[str, Any]] = []
pending_tool_calls: list[dict] = []
for entry in entries:
etype = entry.get("type")
if etype == "message":
role_raw = entry.get("role", "")
from_role = ROLE_MAP.get(role_raw, "gpt")
content = entry.get("content", "")
if not content:
continue
turn: dict[str, Any] = {"from": from_role, "value": content}
# Attach any accumulated tool calls to this gpt turn
if pending_tool_calls and from_role == "gpt":
turn["tool_calls"] = pending_tool_calls
pending_tool_calls = []
turns.append(turn)
elif etype == "tool_call":
tool_name = entry.get("tool", "unknown")
args = entry.get("args", {})
result = entry.get("result", "")
# Record call for the next gpt turn
pending_tool_calls.append({
"name": tool_name,
"arguments": args,
})
# Also emit a tool-result turn immediately after
turns.append({"from": "tool", "value": str(result), "tool": tool_name})
# Discard conversations with < 2 meaningful turns
meaningful = [t for t in turns if t["from"] in ("human", "gpt")]
if len(meaningful) < 2:
return None
return {"conversations": turns}
def load_from_session_logs(logs_dir: Path) -> list[dict[str, Any]]:
"""Load all session JSONL logs and return ShareGPT-formatted conversations."""
log_files = sorted(logs_dir.glob("session_*.jsonl"))
if not log_files:
return []
all_entries: list[dict] = []
for log_file in log_files:
try:
with open(log_file) as f:
for line in f:
line = line.strip()
if line:
try:
all_entries.append(json.loads(line))
except json.JSONDecodeError:
continue
except OSError:
continue
# Sort by timestamp for correct ordering across files
all_entries.sort(key=lambda e: e.get("timestamp", ""))
conversation_groups = _group_into_conversations(all_entries)
results: list[dict[str, Any]] = []
for group in conversation_groups:
conv = _conversation_to_sharegpt(group)
if conv is not None:
results.append(conv)
return results
# ── SQLite fallback reader ───────────────────────────────────────────────────
def load_from_sqlite(db_path: Path) -> list[dict[str, Any]]:
"""Read chat.db and return ShareGPT-formatted conversations."""
if not db_path.exists():
return []
try:
conn = sqlite3.connect(str(db_path))
conn.row_factory = sqlite3.Row
rows = conn.execute(
"SELECT role, content, timestamp FROM chat_messages ORDER BY id"
).fetchall()
conn.close()
except sqlite3.Error:
return []
entries = [
{
"type": "message",
"role": row["role"],
"content": row["content"],
"timestamp": row["timestamp"],
}
for row in rows
]
conversation_groups = _group_into_conversations(entries)
results: list[dict[str, Any]] = []
for group in conversation_groups:
conv = _conversation_to_sharegpt(group)
if conv is not None:
results.append(conv)
return results
# ── Validation ───────────────────────────────────────────────────────────────
def validate_output(output_path: Path) -> dict[str, Any]:
"""Validate the exported JSONL and return stats."""
if not output_path.exists():
return {"error": "Output file not found"}
total = 0
with_tools = 0
turn_counts: list[int] = []
with open(output_path) as f:
for line in f:
line = line.strip()
if not line:
continue
try:
obj = json.loads(line)
except json.JSONDecodeError:
continue
total += 1
turns = obj.get("conversations", [])
turn_counts.append(len(turns))
has_tool = any(
t.get("from") == "tool" or t.get("tool_calls")
for t in turns
)
if has_tool:
with_tools += 1
avg_turns = sum(turn_counts) / len(turn_counts) if turn_counts else 0
return {
"total_conversations": total,
"with_tool_calls": with_tools,
"avg_turns_per_conversation": round(avg_turns, 1),
"output_path": str(output_path),
}
# ── Main ─────────────────────────────────────────────────────────────────────
def build_parser() -> argparse.ArgumentParser:
p = argparse.ArgumentParser(
description="Export Timmy conversation trajectories to ShareGPT JSONL",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
p.add_argument(
"--logs-dir",
type=Path,
default=DEFAULT_LOGS_DIR,
help="Directory containing session_*.jsonl files",
)
p.add_argument(
"--db",
type=Path,
default=DEFAULT_DB_PATH,
help="Path to chat.db (used if no session logs found)",
)
p.add_argument(
"--output",
type=Path,
default=DEFAULT_OUTPUT,
help="Output JSONL file path",
)
p.add_argument(
"--gap-minutes",
type=int,
default=CONVERSATION_GAP_MINUTES,
help="Time gap (minutes) between entries that marks a new conversation",
)
p.add_argument(
"--validate-only",
action="store_true",
help="Skip export; just validate an existing output file",
)
p.add_argument(
"--min-examples",
type=int,
default=0,
help="Exit non-zero if fewer than this many examples are exported",
)
return p
def main(argv: list[str] | None = None) -> int:
args = build_parser().parse_args(argv)
if args.validate_only:
stats = validate_output(args.output)
print(json.dumps(stats, indent=2))
return 0
# ── Load conversations ───────────────────────────────────────────────────
print(f"[1/3] Loading from session logs: {args.logs_dir}")
conversations = load_from_session_logs(args.logs_dir)
if not conversations:
print(f"[1/3] No session logs found — falling back to SQLite: {args.db}")
conversations = load_from_sqlite(args.db)
if not conversations:
print(
"WARNING: No conversation data found.\n"
" • Run the dashboard and have some conversations first.\n"
" • Session logs are written to logs/session_YYYY-MM-DD.jsonl\n"
" • Chat history is stored in data/chat.db",
file=sys.stderr,
)
# Still write empty file so downstream steps don't error on missing file
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text("")
return 0
# ── Write output ─────────────────────────────────────────────────────────
print(f"[2/3] Writing {len(conversations)} conversations → {args.output}")
args.output.parent.mkdir(parents=True, exist_ok=True)
with open(args.output, "w") as f:
for conv in conversations:
f.write(json.dumps(conv) + "\n")
# ── Validate ─────────────────────────────────────────────────────────────
print("[3/3] Validating output…")
stats = validate_output(args.output)
print(json.dumps(stats, indent=2))
if args.min_examples and stats.get("total_conversations", 0) < args.min_examples:
print(
f"ERROR: Only {stats['total_conversations']} examples exported "
f"(need ≥ {args.min_examples})",
file=sys.stderr,
)
return 1
return 0
if __name__ == "__main__":
sys.exit(main())

View File

@@ -0,0 +1,306 @@
"""Unit tests for scripts/export_trajectories.py."""
from __future__ import annotations
import json
import sqlite3
from datetime import datetime, timedelta
from pathlib import Path
import pytest
import scripts.export_trajectories as et
# ── Helpers ──────────────────────────────────────────────────────────────────
def _ts(base: datetime, offset_minutes: int = 0) -> str:
return (base + timedelta(minutes=offset_minutes)).isoformat()
BASE = datetime(2026, 3, 1, 10, 0, 0)
def _make_session_entries(base: datetime = BASE) -> list[dict]:
"""Minimal session log entries: user → tool_call → timmy reply."""
return [
{"type": "message", "role": "user", "content": "list my files", "timestamp": _ts(base, 0)},
{"type": "tool_call", "tool": "shell", "args": {"cmd": "ls"}, "result": "a.py\nb.py", "timestamp": _ts(base, 1)},
{"type": "message", "role": "timmy", "content": "You have two files.", "timestamp": _ts(base, 2)},
]
# ── _group_into_conversations ─────────────────────────────────────────────────
class TestGroupIntoConversations:
def test_empty(self):
assert et._group_into_conversations([]) == []
def test_single_group_no_gap(self):
entries = _make_session_entries()
groups = et._group_into_conversations(entries, gap_minutes=30)
assert len(groups) == 1
assert groups[0] == entries
def test_split_on_large_gap(self):
entries_a = _make_session_entries(BASE)
# Second set starts 60 minutes later
entries_b = _make_session_entries(BASE + timedelta(hours=1))
groups = et._group_into_conversations(entries_a + entries_b, gap_minutes=30)
assert len(groups) == 2
assert len(groups[0]) == 3
assert len(groups[1]) == 3
def test_no_split_within_gap(self):
entries = _make_session_entries()
groups = et._group_into_conversations(entries, gap_minutes=60)
assert len(groups) == 1
def test_entries_without_timestamp(self):
entries = [
{"type": "message", "role": "user", "content": "hello"},
{"type": "message", "role": "timmy", "content": "hi"},
]
groups = et._group_into_conversations(entries, gap_minutes=30)
assert len(groups) == 1
# ── _conversation_to_sharegpt ─────────────────────────────────────────────────
class TestConversationToSharegpt:
def test_basic_exchange(self):
entries = _make_session_entries()
result = et._conversation_to_sharegpt(entries)
assert result is not None
turns = result["conversations"]
human_turns = [t for t in turns if t["from"] == "human"]
gpt_turns = [t for t in turns if t["from"] == "gpt"]
tool_turns = [t for t in turns if t["from"] == "tool"]
assert len(human_turns) == 1
assert len(gpt_turns) == 1
assert len(tool_turns) == 1
def test_tool_calls_attached_to_gpt_turn(self):
entries = [
{"type": "message", "role": "user", "content": "run ls", "timestamp": _ts(BASE, 0)},
{"type": "tool_call", "tool": "shell", "args": {}, "result": "ok", "timestamp": _ts(BASE, 1)},
{"type": "message", "role": "timmy", "content": "done", "timestamp": _ts(BASE, 2)},
]
result = et._conversation_to_sharegpt(entries)
assert result is not None
gpt_turns = [t for t in result["conversations"] if t["from"] == "gpt"]
assert len(gpt_turns) == 1
assert "tool_calls" in gpt_turns[0]
assert gpt_turns[0]["tool_calls"][0]["name"] == "shell"
def test_too_short_returns_none(self):
# Only one meaningful turn → not useful for training
entries = [{"type": "message", "role": "user", "content": "hi", "timestamp": _ts(BASE)}]
assert et._conversation_to_sharegpt(entries) is None
def test_empty_content_skipped(self):
entries = [
{"type": "message", "role": "user", "content": "", "timestamp": _ts(BASE, 0)},
{"type": "message", "role": "timmy", "content": "pong", "timestamp": _ts(BASE, 1)},
]
# Only one non-empty turn → should return None
assert et._conversation_to_sharegpt(entries) is None
def test_role_mapping(self):
entries = [
{"type": "message", "role": "user", "content": "q", "timestamp": _ts(BASE, 0)},
{"type": "message", "role": "assistant", "content": "a", "timestamp": _ts(BASE, 1)},
]
result = et._conversation_to_sharegpt(entries)
assert result is not None
roles = [t["from"] for t in result["conversations"]]
assert "human" in roles
assert "gpt" in roles
def test_decision_entries_ignored(self):
"""Non-message, non-tool entries (decisions, errors) should be skipped."""
entries = _make_session_entries() + [
{"type": "decision", "decision": "do something", "timestamp": _ts(BASE, 10)},
]
result = et._conversation_to_sharegpt(entries)
assert result is not None
assert all(t["from"] != "decision" for t in result["conversations"])
# ── load_from_session_logs ────────────────────────────────────────────────────
class TestLoadFromSessionLogs:
def test_empty_directory(self, tmp_path):
assert et.load_from_session_logs(tmp_path) == []
def test_missing_directory(self, tmp_path):
assert et.load_from_session_logs(tmp_path / "nonexistent") == []
def test_reads_single_log(self, tmp_path):
entries = _make_session_entries()
log = tmp_path / "session_2026-03-01.jsonl"
log.write_text("\n".join(json.dumps(e) for e in entries) + "\n")
result = et.load_from_session_logs(tmp_path)
assert len(result) == 1
assert result[0]["conversations"][0]["from"] == "human"
def test_reads_multiple_logs(self, tmp_path):
for day in range(3):
entries = _make_session_entries(BASE + timedelta(days=day, hours=2 * day))
log = tmp_path / f"session_2026-03-0{day + 1}.jsonl"
log.write_text("\n".join(json.dumps(e) for e in entries) + "\n")
result = et.load_from_session_logs(tmp_path)
# 3 log files, each a separate conversation (days apart)
assert len(result) == 3
def test_skips_malformed_lines(self, tmp_path):
log = tmp_path / "session_2026-03-01.jsonl"
entries = _make_session_entries()
lines = [json.dumps(e) for e in entries]
lines.insert(1, "not valid json{{{")
log.write_text("\n".join(lines) + "\n")
# Should still parse valid entries
result = et.load_from_session_logs(tmp_path)
assert len(result) == 1
# ── load_from_sqlite ──────────────────────────────────────────────────────────
class TestLoadFromSqlite:
def _make_db(self, tmp_path: Path, rows: list[tuple]) -> Path:
db = tmp_path / "chat.db"
conn = sqlite3.connect(str(db))
conn.execute("""
CREATE TABLE IF NOT EXISTS chat_messages (
id INTEGER PRIMARY KEY AUTOINCREMENT,
role TEXT, content TEXT, timestamp TEXT, source TEXT
)
""")
conn.executemany(
"INSERT INTO chat_messages (role, content, timestamp, source) VALUES (?,?,?,?)",
rows,
)
conn.commit()
conn.close()
return db
def test_missing_db(self, tmp_path):
assert et.load_from_sqlite(tmp_path / "missing.db") == []
def test_reads_conversation(self, tmp_path):
rows = [
("user", "hello", _ts(BASE, 0), "browser"),
("agent", "hi there", _ts(BASE, 5), "browser"),
]
db = self._make_db(tmp_path, rows)
result = et.load_from_sqlite(db)
assert len(result) == 1
turns = result[0]["conversations"]
assert turns[0]["from"] == "human"
assert turns[1]["from"] == "gpt"
def test_splits_on_gap(self, tmp_path):
rows = [
("user", "a", _ts(BASE, 0), "browser"),
("agent", "b", _ts(BASE, 5), "browser"),
("user", "c", _ts(BASE, 120), "browser"), # 2h gap
("agent", "d", _ts(BASE, 125), "browser"),
]
db = self._make_db(tmp_path, rows)
result = et.load_from_sqlite(db)
assert len(result) == 2
# ── validate_output ───────────────────────────────────────────────────────────
class TestValidateOutput:
def test_missing_file(self, tmp_path):
stats = et.validate_output(tmp_path / "missing.jsonl")
assert "error" in stats
def test_counts_conversations(self, tmp_path):
out = tmp_path / "out.jsonl"
convs = [
{"conversations": [{"from": "human", "value": "hi"}, {"from": "gpt", "value": "ho"}]},
{"conversations": [{"from": "human", "value": "a"}, {"from": "gpt", "value": "b"}]},
]
out.write_text("\n".join(json.dumps(c) for c in convs) + "\n")
stats = et.validate_output(out)
assert stats["total_conversations"] == 2
assert stats["with_tool_calls"] == 0
def test_counts_tool_calls(self, tmp_path):
out = tmp_path / "out.jsonl"
conv = {"conversations": [
{"from": "human", "value": "run"},
{"from": "gpt", "value": "ok", "tool_calls": [{"name": "shell", "arguments": {}}]},
{"from": "tool", "value": "done", "tool": "shell"},
]}
out.write_text(json.dumps(conv) + "\n")
stats = et.validate_output(out)
assert stats["with_tool_calls"] == 1
# ── CLI (main) ────────────────────────────────────────────────────────────────
class TestMain:
def test_no_data_exits_0(self, tmp_path):
out = tmp_path / "out.jsonl"
code = et.main([
"--logs-dir", str(tmp_path / "logs"),
"--db", str(tmp_path / "missing.db"),
"--output", str(out),
])
assert code == 0
assert out.exists()
def test_exports_from_logs(self, tmp_path):
logs = tmp_path / "logs"
logs.mkdir()
entries = _make_session_entries()
(logs / "session_2026-03-01.jsonl").write_text(
"\n".join(json.dumps(e) for e in entries) + "\n"
)
out = tmp_path / "out.jsonl"
code = et.main([
"--logs-dir", str(logs),
"--db", str(tmp_path / "missing.db"),
"--output", str(out),
])
assert code == 0
lines = [l for l in out.read_text().splitlines() if l.strip()]
assert len(lines) == 1
def test_validate_only(self, tmp_path, capsys):
out = tmp_path / "out.jsonl"
conv = {"conversations": [
{"from": "human", "value": "x"},
{"from": "gpt", "value": "y"},
]}
out.write_text(json.dumps(conv) + "\n")
code = et.main(["--validate-only", "--output", str(out)])
assert code == 0
captured = capsys.readouterr()
stats = json.loads(captured.out)
assert stats["total_conversations"] == 1
def test_min_examples_fails(self, tmp_path):
logs = tmp_path / "logs"
logs.mkdir()
entries = _make_session_entries()
(logs / "session_2026-03-01.jsonl").write_text(
"\n".join(json.dumps(e) for e in entries) + "\n"
)
out = tmp_path / "out.jsonl"
code = et.main([
"--logs-dir", str(logs),
"--db", str(tmp_path / "missing.db"),
"--output", str(out),
"--min-examples", "100",
])
assert code == 1