---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
library_name: peft
pipeline_tag: text-generation
tags:
  - commit-message
  - conventional-commits
  - lora
  - code
  - on-device
---

# Knot Scribe

**On-device model that writes your git commit messages.** Give it a diff, get a
correctly-typed Conventional-Commit message back. Runs 100% on your machine, no
API, no code leaving your device.

Part of the **Knot AI** family, alongside
[**Knot Delta**](https://huggingface.co/pointbreaklab/knot-delta) (which reads a
change and writes a grounded report). Knot Scribe is the model behind
[Knot](https://pointbreaklab.com/knot/)'s hands-free auto-commit, released
standalone here.

> **Renamed:** this repo was formerly `pointbreaklab/knot-ai`. The old URL
> redirects here.

A LoRA fine-tune of **Qwen2.5-Coder-7B-Instruct** (Apache-2.0), trained only on
public commit history. No user data, no private code.

## Benchmark — v5

Sixty hand-reviewed commits (`gold-v1`), balanced across all eleven
conventional-commit types. Head-to-head of the previous release (v4) and the
current model (v5).

| metric | v4 | v5 |
|---|---|---|
| Type accuracy | 48.3% | **60.0%** |
| Scope accuracy | 68.3% | **75.0%** |
| Title word-overlap | 27.2% | 19.1% |
| Exact title | 6.7% | 0.0% |
| Generation errors / 60 | 7 | **4** |

**Type accuracy across releases:** v3 40% → v4 55% → v5 60%. v3/v4 were scored
on an earlier JS/TS-only set; v5 on the harder balanced set above, where v4
re-scores 48.3% — so v5's real gain over v4 is **+11.7 points**.

**Honest read:** v5 lifts type accuracy and scope and nearly halves the format
errors. Title word-overlap and exact-title *fall* because v5 paraphrases: it
writes its own titles, often more descriptive than the single human reference
it is scored against, which exact string-match penalizes. `build` and `revert`
remain hard.

## Use it

**Ollama**

```bash
ollama run pointbreaklab/knot-scribe "$(git diff --staged)"
```

**llama.cpp (base + this LoRA)**

```bash
llama-server --model qwen2.5-coder-7b-instruct-q4_k_m.gguf \
             --lora knot-scribe-lora.gguf
```

Output is JSON `{"title": "...", "body": "..."}` — a correctly-typed
Conventional-Commit subject and body.

## Training

- **Base:** Qwen2.5-Coder-7B-Instruct · Apache-2.0
- **Method:** QLoRA (4-bit), completion-only SFT (loss on the message only)
- **Data:** diff → message pairs mined from permissive open-source commit
  history (JS/TS, Python, Rust, Go)
- **Privacy:** public commit history only — no user data, no private code
- **License:** Apache-2.0 (derivative of Qwen2.5-Coder-7B)