dogmatix

dogmatix β€” 🚧 Coming Soon

dogmatix is a small, local Python coding agent β€” a fine-tune of Google's Gemma-4-E2B-it (β‰ˆ2.3B effective parameters) that runs on a single consumer GPU. It has strong general Python skills and is specialized as a web-application agent: it reads files, edits code, runs the tests, and keeps iterating until they pass.

An initiative and project by Prasad Vittaldev. Weights are being trained now β€” this page will be updated with the released model shortly.

What it does

  • General Python + web specialization: FastAPI, Flask, Django, ORMs (SQLAlchemy / PostgreSQL, Django ORM), and HTML / Jinja templates.
  • Agentic by design: operates in a real coding-agent tool loop (read_file / write_file / list_files / run_tests), doing multi-step read β†’ edit β†’ test β†’ fix.
  • Local & private: runs in 4-bit on a 16 GB GPU β€” no data leaves your machine.
  • Harness-ready: serves behind an Anthropic Messages API or OpenAI Chat Completions endpoint, so it drops into common coding-agent harnesses.

How it's built

A QLoRA fine-tune of gemma-4-E2B-it, distilled from executed, test-verified agent trajectories β€” a stronger teacher model solves sandboxed web-app tasks inside the real tool loop, and only trajectories whose hidden tests pass become training data. Broad Python knowledge comes from curated open Python instruction datasets. The result learns genuine tool use and self-verification, not imagined tool outputs.

πŸ”œ Coming soon (by Prasad Vittaldev)

  • βœ… The released model weights β€” load directly with transformers.
  • πŸ“Š Benchmarks on a held-out web-app coding suite (early runs show a large jump in executed pass@1 and in multi-step tool use vs. the base model).
  • 🧩 End-to-end web-app tasks β€” models + routes + templates + Postgres, together.
  • πŸͺ† Long-task support via task decomposition + context compaction.
  • πŸ§ͺ Recovery-focused training (learning to fix its own failing tests) for higher reliability.
  • πŸ“¦ Ready-to-serve setup for Claude Code / Pi-style harnesses.

License

Apache 2.0, consistent with the base model google/gemma-4-E2B-it.

Contact

dogmatix is an initiative and project by Prasad Vittaldev. For questions or collaboration: prasadvittaldev@gmail.com

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for prasadvittaldev/dogmatix

Finetuned
(280)
this model