dogmatix-GGUF

dogmatix v0.2 — GGUF

GGUF quantisations of dogmatix v0.2, a QLoRA fine-tune of google/gemma-4-E2B-it (~2.3B effective) into a local, Python-specialised web-application coding agent. For llama.cpp / LM Studio / Ollama.

v0.2 supersedes v0.1 (the v0.1 GGUFs are on the v0.1 branch). See the full model card for honest, single-harness benchmarks and limitations.

Files

*Rough llama-bench text-generation on 4 CPU threads; identified as gemma4 E2B, 4.63B.

Q4_K_M is the smallest/fastest and the usual default; Q8_0 is the most faithful to the bf16 weights. All three were verified to load and generate.

What it is for

Small, well-specified Python web work — FastAPI, Flask, SQLAlchemy/SQLModel — through an agent loop (read → edit → test → fix). v0.2 is a sharper web-app specialisation than v0.1: best of the line at end-to-end app-building (pass@1 0.18 base → 0.36 v0.2 on a 12-spec hidden-suite benchmark; monotone across checkpoints, not yet statistically significant), at a measured cost to Django (0.500 vs 0.567 base) and general Python (MBPP 0.700 vs 0.760 base). If you need Django or a general coder as much as app-building, use the v0.1 branch or the base.

Note on tool-calling

dogmatix expects the serving translation layer in the repo for full agentic tool-calling (turn-boundary handling, blind-call guard, client tool-name mapping). Raw GGUF inference works for generation; agentic use through Claude Code / Pi goes through that proxy.

Author

Prasad Vittaldev