Model v4

Model v4 — Diverse Augmentation

Training data now uses real-world package names from PyPI (800+), npm (1300+), crates.io (500+), plus curated docker images, repo names, and system packages. No more myapp hallucinations.

Improvements

• pytohn → python — previously hallucinated myapp:v1.0, now correct
• rm mydir/ → rm -rf mydir/ — previously garbled, now correct
• pip install on PEP 668 — now suggests uvx, pipx, and venv creation
• docker ps permission — now suggests both sudo and usermod -aG docker
• All v3 fixes retained (clean EOS stopping, multi-alt where appropriate)

Stats

• 60K training examples with 2600+ unique package/project names
• No single placeholder exceeds 0.1% of training data (was 7% for myapp)
• Train loss: 0.099, eval loss: 0.068

Model v3

Model v3 — EOS Fix

Training improvements:

• Added EOS token to completions — model now learns when to stop generating, eliminating hallucinated multi-line garbage
• Aligned prompt format — leading space moved from completion into prompt (OP: instead of OP: + )

Results:

• Single-op cases (typos, sudo) produce clean single results
• Multi-alt cases (ambiguous errors) still produce valid alternatives
• No more nonsensical hallucinations after the first valid op

Known limitations (targeted for v4):

• Augmentation data has low diversity — myapp placeholder overrepresented (~7% of training data), causing hallucinations on unfamiliar commands

Model v2

update model checksums for Model v2

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Model v1

To install, follow the instructions in the README. This release exists to provide model downloads to the shit CLI.

Initial release — fine-tuned Gemma 3 270M (Q4_K_M) for command correction.