Khasi-OCR-v1

This is a fine-tuned version of DeepSeek-OCR-2 specifically optimized for high-accuracy Optical Character Recognition (OCR) of the Khasi language.

It was trained on a custom dataset of Khasi news articles, official documents, and literature to overcome the common hallucination and repetition issues (such as infinite loops) found in base multimodal models when handling low-resource languages.

Model Highlights

• Language Support: Native Khasi (using Latin script with special characters like ï and ñ).
• Task: Specialized for "Free OCR" (literal document transcription).
• Base Model: DeepSeek-OCR-2.
• Accuracy: Significantly lower Character Error Rate (CER) & Word Error Rate (WER)

Usage

To use this model on Kaggle or a local GPU, ensure you have transformers, accelerate, and bitsandbytes installed.

from transformers import AutoModel, AutoTokenizer, BitsAndBytesConfig
import torch

model_id = "toiar/deepseek-ocr-2-khasi-v1"

# Load the model in 4-bit to save memory
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    quantization_config=bnb_config,
    device_map="auto"            
)

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

# Set your image path and prompt
image_path = "your_khasi_image.jpg" 
prompt = "<image>\nFree OCR."

# Run the transcription
res = model.infer(
    tokenizer,
    prompt=prompt,
    image_file=image_path,
    output_path="./outputs",
    save_results=False
)

# View results
if isinstance(res, dict) and "text" in res:
    print(res["text"])
else:
    print(res)