seshat-commit

Conventional Commits message generator in Brazilian Portuguese (PT-BR), fine-tuned from Qwen 2.5 Coder 7B.

Receive a git diff, return a single-line commit message in Conventional Commits format. Trained on 4869 real commits from the author’s repositories with strict quality filters. Designed to be used inside Seshat, an automated commit CLI.

Model page: https://ollama.com/juniormartinxo/seshat-commit

Quick start

ollama pull juniormartinxo/seshat-commit

# basic usage:
git diff --cached | ollama run juniormartinxo/seshat-commit

# with Seshat (recommended):
seshat config --provider ollama --model juniormartinxo/seshat-commit
seshat commit --yes

Example:

input  → diff --git a/src/foo.rs b/src/foo.rs
         +log::info!("starting bar");

output → feat(foo): adicionar log de início na função bar

Specs

What it does well

• Generates Conventional Commits in PT-BR (feat, fix, chore, docs, refactor, perf, test, build, ci, style, revert)
• Infers correct scope from file paths (feat(api), fix(rtk), chore(release))
• Mixed PT-BR/EN naturally for technical terms (bump version, endpoint)
• Handles Rust, Python, TypeScript, JavaScript, Go, Shell, Markdown
• Validates against Conventional Commits spec on ~95–98% of typical diffs

What it doesn’t do

• Long-form text (release notes, PR descriptions, code generation)
• Conversational chat — it’s specialized for diff → commit
• English-only output — output is PT-BR by default

Recommended parameters

The included Modelfile sets:

PARAMETER temperature 0.2
PARAMETER top_p 0.9
PARAMETER num_ctx 8192
PARAMETER repeat_penalty 1.05

SYSTEM "Você é um gerador de mensagens de commit no padrão Conventional Commits.
Receba um git diff e responda apenas com a mensagem de commit, sem explicação.
Use PT-BR no corpo quando aplicável.
Tipo válido: feat, fix, chore, docs, refactor, perf, test, build, ci, style, revert."

Lower temperature (0.0–0.3) for production, higher (0.5–0.8) if you want more variation.

Training data

Extracted from public + private repositories of the author (~14k commits → 5k after filtering). Filters applied:

• Only Conventional Commits format (regex ^(feat|fix|chore|...)(scope)?(!)?: .+)
• Removed: WIP, Revert, fixup!, squash!, [skip ci], bot authors, whitespace-only diffs, generic subjects (update, fix, tweak…), placeholders (<TODO>, XXX), URL-only subjects, filename-only subjects
• Deduplicated by subject + diff hash
• Capped per type to balance distribution
• Languages observed: Rust (40%), TypeScript (25%), Python (15%), JS/Go/Shell (20%)

Hardware requirements

Runs comfortably on a single consumer GPU. CPU inference works but slow (~10–20 tokens/s on modern x86).

License

The base model (Qwen 2.5 Coder) is released under the Qwen Research License. This fine-tune inherits the same license.

The fine-tuned LoRA adapter and dataset extraction pipeline are MIT.

Source / pipeline

Full reproducible training pipeline at: https://github.com/juniormartinxo/seshat (see scripts/dataset/ and scripts/training/).

Re-train on your own commit history in ~1 hour:

make junior              # extract + normalize your commits
python train.py          # LoRA fine-tune
ollama create my-model -f Modelfile

Limitations & known issues

• Subjects > 100 chars sometimes get cut mid-sentence
• Rare diffs (heavy renames, binary deletions) may produce English fallback
• Bias towards author’s style: terse, lowercase-after-colon, occasional EN/PT mix
• No multi-paragraph commit bodies (always single-line subject)

Use with Seshat

Seshat is a CLI that automates the commit workflow:

seshat init                                              # configure project
seshat config --provider ollama --model juniormartinxo/seshat-commit
git add .
seshat commit --yes                                      # generate + commit
seshat flow 5 --yes                                      # batch-commit modified files
seshat bench agents --agents ollama --show-samples 3     # benchmark vs Codex/Claude

Tags

conventional-commits · git · commit-message · portuguese · pt-br · qwen · code · lora · fine-tuned