README.md

---
base_model: unsloth/DeepSeek-OCR-2
library_name: transformers
pipeline_tag: image-to-text
language:
- th
tags:
- ocr
- thai
- thai-ocr
- handwriting
- unsloth
- peft
- lora
- merged
- deepseek-ocr2
---

# DeepSeek-OCR-2 Thai OCR (Trained Model)

This repository contains the final Thai OCR model weights that were trained from `unsloth/DeepSeek-OCR-2` and exported as a full merged model for inference.

## Model Summary

- Task: Thai OCR text extraction (especially handwriting)
- Base model: `unsloth/DeepSeek-OCR-2`
- Fine-tuning method: LoRA
- Final artifact in this repo: merged full model (not adapter-only)
- Prompt format used during training/inference:
  - `<image>\nOCR this image and output the Thai text.`

## Training Data

- Dataset: `iapp/thai_handwriting_dataset`
- Main training portions used:
  - `train[0:10150]`
  - `train[10150:13600]` (continuation phase)

## Training Configuration

LoRA configuration:
- `target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]`
- `r = 16`
- `lora_alpha = 16`
- `lora_dropout = 0`
- `bias = "none"`

Main optimization setup:
- `per_device_train_batch_size = 1`
- `gradient_accumulation_steps = 4`
- `learning_rate = 2e-4`
- `optim = "adamw_8bit"`
- `lr_scheduler_type = "linear"`

Vision preprocessing setup:
- `base_size = 1024`
- `image_size = 768`
- `crop_mode = True`
- `auto_resize = True`
- `max_dynamic_crops = 6`

## Final Model Weights in This Repo

The model published to Hugging Face is a merged model exported to safetensors shards.

Expected key files:
- `model-00001-of-00002.safetensors`
- `model-00002-of-00002.safetensors`
- `model.safetensors.index.json`
- `config.json`
- tokenizer and processor files

## Quick Inference

```python
from transformers import AutoModel
from unsloth import FastVisionModel

model, tokenizer = FastVisionModel.from_pretrained(
    model_name="phoritus/deepseek-ocr2-thai",
    auto_model=AutoModel,
    trust_remote_code=True,
    load_in_4bit=False,
    device_map="auto",
)

result = model.infer(
    tokenizer,
    prompt="<image>\nOCR this image and output the Thai text.",
    image_file="sample_image.jpg",
    output_path="output_results",
    base_size=1024,
    image_size=768,
    crop_mode=True,
    save_results=True,
    test_compress=False,
)

print(result)
```

## Evaluation Snapshot (Infer Path, Random 10)

JiWER OCR test summary:
- `samples`: 10
- `wer`: 1.0
- `cer`: 0.648695652173913
- `mer`: 1.0
- `wil`: 1.0
- `wip`: 0.0
- `hits`: 0
- `substitutions`: 12
- `deletions`: 2
- `insertions`: 0

Per-sample results:

| order | source_idx | sample_wer | sample_cer |
| --- | --- | --- | --- |
| 1 | 440 | 1.0 | 0.8823529411764706 |
| 2 | 996 | 1.0 | 0.17543859649122806 |
| 3 | 687 | 1.0 | 0.7 |
| 4 | 124 | 1.0 | 0.7065217391304348 |
| 5 | 408 | 1.0 | 0.5074626865671642 |
| 6 | 46 | 1.0 | 0.36 |
| 7 | 404 | 1.0 | 0.7659574468085106 |
| 8 | 254 | 1.0 | 0.7941176470588235 |
| 9 | 666 | 1.0 | 0.8490566037735849 |
| 10 | 978 | 1.0 | 0.8823529411764706 |

## Limitations

- Best quality is expected on Thai receipt/handwriting-like inputs similar to training data.
- Performance can degrade on low-quality, rotated, or heavily noisy images.
- For production use, validate with your own evaluation set and prompt constraints.

## Acknowledgements

- Base model: `unsloth/DeepSeek-OCR-2`
- Libraries: Unsloth, Transformers, PEFT
- Dataset: `iapp/thai_handwriting_dataset`