Type what you want. Get the shell command.
ShellVibe is a family of fine-tuned language models that translate plain English into shell commands — running fully local, with no API keys or internet connection required.
Most developers waste time Googling arcane flags or piping through man pages. ShellVibe puts a fine-tuned model directly in your terminal: describe your intent, get the exact command.
- Fully local — no API keys, no data sent anywhere
- Three model sizes — choose speed vs. accuracy for your hardware
- GGUF format — runs efficiently on CPU and Apple Silicon via llama.cpp
- Trained on real shell data — sourced from tldr-pages, covering macOS and Linux
⚠️ Always verify generated commands before running them. ShellVibe can make mistakes — especially with complex flags or destructive operations (rm,chmod,dd, etc.). Treat the output as a suggestion, not a guarantee.
All models are fine-tuned from Qwen2.5-Coder-Instruct and trained on an NVIDIA A100 (bf16).
| Model | Base | Size | W&B Training Logs |
|---|---|---|---|
| ShellVibe-0.5B | Qwen2.5-Coder-0.5B-Instruct | ~500MB | view run |
| ShellVibe-1.5B | Qwen2.5-Coder-1.5B-Instruct | ~1.5GB | view run |
| ShellVibe-3B | Qwen2.5-Coder-3B-Instruct | ~3GB | view run |
| 0.5B | 1.5B | 3B |
 |
 |
 |
brew install uv
uv syncDownload the gguf-models/ folder from Google Drive and place it at the root of the repo.
gguf-models/
├── qwen2.5-0.5b-inst-q8_0.gguf
├── qwen2.5-1.5b-inst-q8_0.gguf
└── qwen2.5-3b-inst-q8_0.gguf
Pick a model size and run — backend is auto-detected (Metal on macOS, CPU otherwise):
make run-0.5b # fastest, lightest
make run-1.5b # balanced
make run-3b # most accurateTokens/second is displayed after each response.
The models were fine-tuned using supervised fine-tuning (SFT) on shell command data derived from tldr-pages. Training was done on an NVIDIA A100 with bf16 precision.
make preprocess-tldr # parse TLDR markdown pages → CSVmake trainCheckpoints are saved automatically on best validation loss and best edit-distance score. Training metrics are logged to Weights & Biases.
GPL
