DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA-v0.1

A small LoRA adapter for Gemma 4 26B-A4B MLX 4-bit, focused on Thai OCR-style prompts.

The goal is simple: compare against the unmodified Gemma base and test whether a tiny LoRA can make Thai OCR output shorter and cleaner. This is not a claim against dedicated OCR systems.

What Changed

The adapter mainly improves output discipline. It reduces verbose preambles and makes the model answer OCR prompts more directly.

Eval set	Metric	Gemma base	+ this LoRA	Change
ThaiOCRBench subset, n=299	CER	3.130	1.792	-43%
ThaiOCRBench subset, n=299	DW-light score	0.5215	0.5382	+0.017
openthaigpt/thai-ocr-evaluation, n=104	CER	2.350	1.730	-26%
openthaigpt/thai-ocr-evaluation, n=104	DW-light score	0.9009	0.9079	+0.007

The score changes are small and within noise. The useful change is CER: the adapter makes the base model less wordy on Thai OCR tasks.

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-v0.1",
)

img = Image.open("doc.jpg").convert("RGB")
messages = [{"role": "user", "content": "ถอดข้อความในภาพ"}]
prompt = apply_chat_template(processor, model.config, messages, num_images=1)

out = generate(
    model,
    processor,
    image=img,
    prompt=prompt,
    max_tokens=512,
    temp=0.0,
    verbose=False,
)
print(out.text)

Tested on Apple Silicon with 24 GB+ unified memory. The development run used an M4 Max 64 GB.

Method

This is a rank-4 LoRA trained for 800 iterations with mlx-vlm.

Key choices:

• Rank 4 LoRA: about 5M trainable parameters, roughly 0.077% of the base.
• Very low LR: 1e-6. Higher learning rates caused output collapse in smoke runs.
• Retention prompts: 100 iAPP handwriting images with 4 Thai prompt variants each.
• Chunked training: eval every 200 iterations; selected the 800-iteration checkpoint.
• Auto-stop gates: stopped before the 1000-iteration checkpoint because document classification regressed.
• Contamination check: 1 text-overlap sample was excluded from the iAPP train pool.

Training data:

source: iAPP Thai Handwriting Dataset
license: Apache 2.0
train size: 400 records
recipe: 100 unique images x 4 prompt variants

Task Notes

Compared with the Gemma base on ThaiOCRBench:

Area	Direction
Text recognition	improved
Table parsing	improved
Diagram VQA	improved
KIE / KIM / reasoning tasks	mostly preserved
Full-page OCR	regressed
Document classification	small regression from a base ceiling

Use the base model instead if full-page OCR is the main task.

Evaluation Notes

ThaiOCRBench numbers use a 299-sample stratified subset with seed 42.

DW-light score is max(substring_match, ANLS). It is a lightweight local scoring protocol, not the full ThaiOCRBench paper protocol. Do not compare these numbers directly to paper leaderboard numbers.

The neutral set is openthaigpt/thai-ocr-evaluation with 104 samples. It was used only for evaluation.

Limitations

• Research preview, not production-certified OCR.
• The adapter improves CER, not broad accuracy.
• Full-page OCR regresses versus the base.
• Mac / MLX only. No GGUF build in this release.
• Not tested on a DW-collected private hold-out set yet.
• The adapter does not add new language ability; it sharpens Thai OCR-style behavior already present in the base.

Files

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

License And Credits

This LoRA adapter is Apache 2.0.

Built on:

• Gemma 4 26B-A4B
• mlx-community/gemma-4-26b-a4b-it-4bit
• iAPP Thai Handwriting Dataset
• ThaiOCRBench, evaluation only
• openthaigpt/thai-ocr-evaluation, evaluation only
• mlx-vlm

Citation

@misc{dw_khotaevl_ocr_2026,
  title  = {DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA v0.1},
  author = {Deaw},
  year   = {2026},
  url    = {https://huggingface.co/commandeaw/DW-KhoTaeVL-OCR-Pro-26B-MLX-LoRA-v0.1}
}