Robb Doering
All work

myModel

Model serving, OCR, and tuning tools for my own GPU and the cloud.

Source on GitLab Python · MPL-2.0 · v0.6.0
Two stacked bars of Python line counts measured with wc -l over the tracked .py files of mymodel. The top bar splits the four subpackages, 27,274 lines in all (the top-level modules serve, ternary, web and common add 834 more and are not drawn): ocr has 22,998 lines (84 percent, cyan), ocr_fixes 2,996 (11 percent), catalogue 908 (3 percent), and tune 372 (1 percent). The bottom bar zooms into ocr: quality has 13,502 lines (59 percent, cyan), dashboard 1,621 (7 percent, green), and everything else 7,875 (34 percent, grey). 4 subpackages 27,274 lines ocr_fixes/ 2,996 · 11% catalogue/ 908 · 3% tune/ 372 · 1% ocr/ 22,998 · 84% ocr/ 22,998 lines quality/ 13,502 · 59% other 7,875 · 34% dashboard/ 1,621 · 7%

myModel collects the tools around running language models on my 4090 & the cloud: catalogue-driven vLLM serving behind a small management API, a llama.cpp session service, OCR with Marker, and QLoRA fine-tuning. Lately I've been using it almost exclusively for OCR. This one is less polished and almost entirely agent-driven at this point, but it's still reliable and relevant.

The pieces need very different runtime environments, so they share source and schemas but install separately. The Epstein Index builds on it.