[claude] Export conversation trajectories to ShareGPT JSONL (#1102) #1121

claude · 2026-03-23T18:20:36Z

claude commented

2026-03-23 18:20:36 +00:00

What this does

Implements AutoLoRA Step 3 of 7: exports Timmy conversation trajectories to ShareGPT-compatible JSONL for Hermes 4 LoRA fine-tuning.

New files

scripts/export_trajectories.py — main export script
tests/scripts/test_export_trajectories.py — 26 unit tests (all passing)

Usage

# Export everything to ~/timmy-training-data.jsonl
python scripts/export_trajectories.py

# Custom paths
python scripts/export_trajectories.py --logs-dir ./logs --output ~/training.jsonl

# Validate existing output
python scripts/export_trajectories.py --validate-only --output ~/timmy-training-data.jsonl

# Fail if fewer than 100 examples
python scripts/export_trajectories.py --min-examples 100

Data sources (priority order)

logs/session_*.jsonl — session logs with full tool call info (preferred)
data/chat.db — SQLite chat history fallback

Output format (ShareGPT)

{"conversations": [
  {"from": "human", "value": "list my files"},
  {"from": "gpt", "value": "done", "tool_calls": [{"name": "shell", "arguments": {}}]},
  {"from": "tool", "value": "a.py\nb.py", "tool": "shell"}
]}

Conversations are split on 30-minute gaps between entries. Conversations with fewer than 2 meaningful turns are discarded.

Fixes #1102 ## What this does Implements AutoLoRA Step 3 of 7: exports Timmy conversation trajectories to ShareGPT-compatible JSONL for Hermes 4 LoRA fine-tuning. ## New files - `scripts/export_trajectories.py` — main export script - `tests/scripts/test_export_trajectories.py` — 26 unit tests (all passing) ## Usage ```bash # Export everything to ~/timmy-training-data.jsonl python scripts/export_trajectories.py # Custom paths python scripts/export_trajectories.py --logs-dir ./logs --output ~/training.jsonl # Validate existing output python scripts/export_trajectories.py --validate-only --output ~/timmy-training-data.jsonl # Fail if fewer than 100 examples python scripts/export_trajectories.py --min-examples 100 ``` ## Data sources (priority order) 1. `logs/session_*.jsonl` — session logs with full tool call info (preferred) 2. `data/chat.db` — SQLite chat history fallback ## Output format (ShareGPT) ```json {"conversations": [ {"from": "human", "value": "list my files"}, {"from": "gpt", "value": "done", "tool_calls": [{"name": "shell", "arguments": {}}]}, {"from": "tool", "value": "a.py\nb.py", "tool": "shell"} ]} ``` Conversations are split on 30-minute gaps between entries. Conversations with fewer than 2 meaningful turns are discarded.

claude added 1 commit 2026-03-23 18:20:37 +00:00

feat: export conversation trajectories to ShareGPT JSONL for LoRA fine-tuning

Tests / test (pull_request) Has been skipped

Details

Tests / lint (pull_request) Failing after 16s

Details

a2f8989c39

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>

claude referenced this pull request

2026-03-23 18:21:04 +00:00

[AutoLoRA P1] Export Claude Conversation Trajectories #1102

Rockachopa commented

2026-03-23 18:21:43 +00:00

LGTM — clean feature with 26 tests. Has merge conflicts. Please rebase on main and force-push, then I will merge.

Rockachopa referenced this pull request

2026-03-23 18:21:44 +00:00

[claude] Fix TemplateResponse signature for Starlette new API (#1114) #1120

Timmy commented

2026-03-23 18:27:47 +00:00

Closing: PR has merge conflicts with main that cannot be auto-rebased. The underlying issue remains open for re-implementation.