xav-ie/scribe-cleanup-lora

scribe cleanup v1 (LoRA adapter)

You probably want a ready-made version instead: the full model, the Mac version or the llama.cpp version. What the model does, its results, limits and how it was built are on the main card.

This is the trained add-on itself (LoRA rank 16, alpha 32, 93 MB) for google/gemma-4-E2B-it. It shows exactly what training changed, and lets you build your own versions.

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoTokenizer
tok = AutoTokenizer.from_pretrained("google/gemma-4-E2B-it")
base = AutoModelForImageTextToText.from_pretrained("google/gemma-4-E2B-it", dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "xav-ie/scribe-cleanup-lora").merge_and_unload()

Then prompt it as the main card's "How to use" shows, with prompt.txt, and check the result with guard.py.

Merge it into the full-precision base, then compress. Merged into an already compressed 4-bit Gemma, most of the adapter rounds away: self-corrections are silently lost while ordinary cleanup still looks fine. All released compressed versions were made from the full-precision merge, as above.

License: Apache-2.0 (LICENSE), built on google/gemma-4-E2B-it (Apache-2.0).