[claude] LoRA trajectory export and fine-tune launcher (#1103) #1117

claude · 2026-03-23T18:15:29Z

claude commented

2026-03-23 18:15:29 +00:00

What this adds

AutoLoRA Step 4 of 7 — tooling to convert Timmy session logs into LoRA training data and launch fine-tuning on Apple Silicon.

`scripts/export_trajectories.py`

Converts session_*.jsonl logs (written by SessionLogger) into mlx-lm ChatML training examples:

Groups session entries into (user, assistant) turn pairs
Embeds tool calls as <tool_call> XML blocks in assistant turns (Hermes 4 format)
Filters trivially short responses
Writes ~/timmy-training-data.jsonl

`scripts/lora_finetune.py`

Wrapper around mlx_lm.lora with project-tuned defaults:

Pre-flight checks: Apple Silicon, mlx-lm installed, model path, train data present
--dry-run prints the exact command without executing
--test runs inference with the trained adapter
--fuse merges adapter into base model for Ollama import
Default: batch=1, lora-layers=16, iters=1000, lr=1e-5 (conservative for 36 GB M3 Max)

`tests/scripts/test_export_trajectories.py`

20 unit tests: all passing.

Workflow

# 1. Generate training data
python scripts/export_trajectories.py --verbose

# 2. Prepare training dir
mkdir -p ~/timmy-lora-training
cp ~/timmy-training-data.jsonl ~/timmy-lora-training/train.jsonl

# 3. Fine-tune (~2-8h on M3 Max)
python scripts/lora_finetune.py

# 4. Test
python scripts/lora_finetune.py --test

# 5. Fuse + import to Ollama
python scripts/lora_finetune.py --fuse

Fixes #1103 ## What this adds AutoLoRA Step 4 of 7 — tooling to convert Timmy session logs into LoRA training data and launch fine-tuning on Apple Silicon. ### `scripts/export_trajectories.py` Converts `session_*.jsonl` logs (written by `SessionLogger`) into `mlx-lm` ChatML training examples: - Groups session entries into `(user, assistant)` turn pairs - Embeds tool calls as `<tool_call>` XML blocks in assistant turns (Hermes 4 format) - Filters trivially short responses - Writes `~/timmy-training-data.jsonl` ### `scripts/lora_finetune.py` Wrapper around `mlx_lm.lora` with project-tuned defaults: - Pre-flight checks: Apple Silicon, mlx-lm installed, model path, train data present - `--dry-run` prints the exact command without executing - `--test` runs inference with the trained adapter - `--fuse` merges adapter into base model for Ollama import - Default: batch=1, lora-layers=16, iters=1000, lr=1e-5 (conservative for 36 GB M3 Max) ### `tests/scripts/test_export_trajectories.py` 20 unit tests: all passing. ## Workflow ```bash # 1. Generate training data python scripts/export_trajectories.py --verbose # 2. Prepare training dir mkdir -p ~/timmy-lora-training cp ~/timmy-training-data.jsonl ~/timmy-lora-training/train.jsonl # 3. Fine-tune (~2-8h on M3 Max) python scripts/lora_finetune.py # 4. Test python scripts/lora_finetune.py --test # 5. Fuse + import to Ollama python scripts/lora_finetune.py --fuse ```

claude added 1 commit 2026-03-23 18:15:30 +00:00

feat: add LoRA trajectory export and fine-tune launcher scripts (AutoLoRA Step 4)

Tests / lint (pull_request) Failing after 16s

Details

Tests / test (pull_request) Has been skipped

Details

97c0d06204

Adds the tooling for Step 4 of the AutoLoRA Sovereignty Loop:
- scripts/export_trajectories.py: converts session JSONL logs into
  mlx-lm ChatML training data, embedding tool calls as <tool_call> blocks
  in assistant turns so the fine-tuned model learns structured tool use
- scripts/lora_finetune.py: wraps mlx_lm.lora/generate/fuse with
  project-specific defaults, pre-flight checks, and dry-run support
- tests/scripts/test_export_trajectories.py: 20 unit tests covering
  entry loading, grouping, tool-call formatting, and full export pipeline

Workflow:
  python scripts/export_trajectories.py --verbose
  cp ~/timmy-training-data.jsonl ~/timmy-lora-training/train.jsonl
  python scripts/lora_finetune.py
  python scripts/lora_finetune.py --test
  python scripts/lora_finetune.py --fuse

Fixes #1103

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

claude referenced this pull request

2026-03-23 18:15:38 +00:00

[AutoLoRA P1] LoRA Fine-Tune Hermes 4 on Exported Trajectories #1103

claude merged commit 32a5b092d0 into main

2026-03-23 18:15:46 +00:00

claude deleted branch claude/issue-1103

2026-03-23 18:15:46 +00:00

claude referenced this issue from a commit

2026-03-23 18:15:48 +00:00

[claude] LoRA trajectory export and fine-tune launcher (#1103) (#1117)

Sign in to join this conversation.