DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA-v0.2

A small LoRA for Gemma 4 26B-A4B MLX 4-bit, Thai OCR.

Sibling to v0.1, not a replacement. v0.1 was sharp on short OCR but broke on full pages. v0.2 trades some of that sharpness back for full-page recovery. Same recipe, different training prompts.

Numbers

Overall score is tied with v0.1. Real movement is per-task: Full-page +0.13, Text recognition -0.04. Pick v0.1 for short OCR, v0.2 for documents.

Quickstart

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from PIL import Image

model, processor = load(
    "mlx-community/gemma-4-26b-a4b-it-4bit",
    adapter_path="commandeaw/DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA-v0.2",
)
img = Image.open("doc.jpg").convert("RGB")
messages = [{"role": "user", "content": "ถอดข้อความทั้งหมดในเอกสารนี้"}]
prompt = apply_chat_template(processor, model.config, messages, num_images=1)
print(generate(model, processor, image=img, prompt=prompt, max_tokens=512, temp=0.0).text)

Tested on Apple Silicon 24 GB+. Dev run on M4 Max 64 GB.

Method

Same as v0.1: rank 4 LoRA, lr 1e-6, batch 1, grad_accum 16, mlx-vlm. ~5M trainable params (0.077% of base). Only the data changed.

v0.2 has 1200 records (100 iAPP images × 12 prompts). For 10 of the 12 prompts (long-form OCR, doc classify, KIE), the gold answer is the base model's own output — self-distillation, to anchor the LoRA to base behavior on prompts v0.1 broke. The other 2 (short_transcribe, short_read) use iAPP ground truth.

Trained 400 iters. Auto-stop fired at 600 when document classification collapsed.

Task Notes

Eval Notes

299-sample stratified subset of ThaiOCRBench (seed 42), same set v0.1 used. DW-light = max(substring_match, ANLS) — lightweight local scoring, not the paper protocol. Neutral set is openthaigpt/thai-ocr-evaluation, n=104.

Limitations

• Research preview, not production OCR.
• Short-text CER is worse than v0.1. Use v0.1 if that's all you need.
• DocCls hasn't recovered to base (0.926 vs 1.000), same gap as v0.1.
• Mac / MLX only.
• Shifts existing Thai OCR behavior; doesn't add new ability.

Files

adapters.safetensors
adapter_config.json
train_manifest.jsonl
contamination_manifest_full.json
NOTICE
FROZEN.md

License

Apache 2.0. Built on Gemma 4 26B-A4B (Apache 2.0), iAPP Thai Handwriting Dataset (Apache 2.0), mlx-vlm. Evaluated on ThaiOCRBench and openthaigpt/thai-ocr-evaluation.

Citation

@misc{dw_khotaevl_ocr_v02_2026,
  title  = {DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA v0.2},
  author = {Deaw},
  year   = {2026},
  url    = {https://huggingface.co/commandeaw/DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA-v0.2}
}