em-router

Skill-routing and commit-message model fine-tuned for the em developer CLI tool.

Given a free-text developer request → outputs the single best skill slug (sdd-feature, ws-commit, gitnexus-exploring, …). Given a unified git diff → outputs a one-line Conventional Commit message.

Model Details

Field	Value
Developed by	Travis Dinh (ninosh)
Base model	Qwen/Qwen2.5-1.5B-Instruct
Model type	LoRA fine-tune → GGUF Q4_K_M
Language	English
License	Apache-2.0
Tasks	Skill routing (classification), commit message generation

Training

Hyperparameter	Value
Framework	Unsloth + TRL SFTTrainer
Hardware	Kaggle T4 (fp16)
LoRA rank / alpha	16 / 32
Epochs	3
Effective batch size	16
Learning rate	2e-4
Final eval loss	1.369

Training Data

Dataset	Rows	Task
ninosh/em-commit-msg-train	150	Conventional commit from diff
ninosh/em-intent-train	240	Skill slug from developer request

Both use OpenAI chat format (system / user / assistant), combined and split 90/10 train/val.

Files

File	Size	Description
em-router-Q4_K_M.gguf	~986 MB	4-bit quantized — use with llama.cpp or Ollama
adapter_model.safetensors	~71 MB	Raw LoRA adapter — use with PEFT/Unsloth for further fine-tuning
tokenizer.json	11 MB	Qwen2.5 tokenizer

Usage

With em CLI (recommended)

# Download once (~986 MB)
em model pull

# Generate a commit message from staged diff
git diff HEAD | em model run --task commit-msg

# Route a request to the best skill
em model run --task intent --prompt "I need to review my recent changes"

With llama.cpp

huggingface-cli download ninosh/em-router em-router-Q4_K_M.gguf
llama-cli -m em-router-Q4_K_M.gguf --chat-template chatml \
  -p "You are a skill router. Output only one skill slug. No explanation.\nUser: review my changes"

With Ollama

ollama run hf.co/ninosh/em-router

Skill Set Version

Trained against skill set hash 86f22cbf. Run em model status to check whether a newer version is available.

Limitations

• Covers the em skill library (~80 skills). Out-of-scope requests return the closest match.
• Commit messages are single-line only; no body/footer generation.
• Context window 2048 tokens; very large diffs are truncated.