pointbreaklab/ knot-scribe:latest

6 days ago

tools 
ollama run pointbreaklab/knot-scribe

Details

6 days ago

8ca7308c465a · 5.0GB ·

model
qwen2
·
7.62B
·
Q4_K_M
template
{{- if .Suffix }}<|fim_prefix|>{{ .Prompt }}<|fim_suffix|>{{ .Suffix }}<|fim_middle|> {{- else if .M
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
adapter
qwen2
·
161M
system
You are Knot AI. Your name is always Knot AI. Recognize your name case-insensitively and with or wit
params
{ "num_ctx": 32768, "temperature": 0.15, "top_p": 0.8 }

Readme

Knot Scribe

Writes your git commit messages, on your machine. Give it a diff, get back a correctly-typed Conventional-Commit message. No API, no cloud, no code leaving your device.

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

Output is JSON {"title": "...", "body": "..."} — an imperative type(scope): summary title and an optional body.

Part of the Knot AI family with pointbreaklab/knot-delta (which explains what changed). Knot Scribe is the model behind Knot’s hands-free auto-commit.

What the fine-tune buys — vs the base model

Knot Scribe is a LoRA fine-tune of Qwen2.5-Coder-7B. Here it is against that untouched base model, scored on the same benchmark: 60 hand-reviewed commits (gold-v1), balanced across all 11 conventional-commit types.

Type accuracy — Knot Scribe vs its base model (gold-v1)

Qwen2.5-Coder-7B  (base)  ████████                          20%
Knot Scribe v5            ████████████████████████          60%
                          └ 3x the type accuracy from fine-tuning the same base
metric Qwen2.5-Coder-7B (base) Knot Scribe v5
Type accuracy 20.0% 60.0%
Scope accuracy 35.0% 75.0%
Malformed replies / 60 11 4

The base model returns unstructured prose it can’t reliably parse; the fine-tune makes the output a dependable JSON commit message and picks the right type 3× as often.

Progress across releases (same gold-v1 set)

Type accuracy

base   ████████                          20%
v4     ███████████████████               48%
v5     ████████████████████████          60%

v4 → v5, full metrics (gold-v1, 60 commits)

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

By commit type (correct / n)

type v4 v5 type v4 v5
chore 5⁄6 6⁄6 docs 4⁄6 5⁄6
ci 5⁄6 6⁄6 perf 2⁄6 2⁄6
feat 4⁄6 4⁄6 build 0/5 0/5
fix 5⁄6 4⁄6 refactor 0/6 2⁄6
test 4⁄5 4⁄5 revert 0/2 0/2

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’s scored against, which exact string-match penalizes. build and revert stay hard.

Model

  • Base: Qwen2.5-Coder-7B-Instruct · Apache-2.0
  • Method: QLoRA (4-bit), completion-only SFT on public commit history
  • Privacy: trained on public commits only — no user data, no private code
  • Also on: Hugging Face
  • License: Apache-2.0