Update README.md

README.md

@@ -6,32 +6,49 @@ tags:
 - handwriting-recognition
 - vision-language
 - fine-tuned
+- vision
 datasets:
 - iapp/thai_handwriting_dataset
 language:
 - th
+pipeline_tag: image-to-text
 ---
 
-# Thai Handwriting Recognition Model
+# Thai Handwriting Recognition Vision-Language Model
 
-This is a LoRA-based fine-tuned model based on Llama-3.2-11B-Vision-Instruct for Thai handwriting recognition.
+A LoRA-adapted vision-language model based on Llama-3.2-11B-Vision-Instruct that transcribes Thai handwritten text from images.
 
-## Model Details
+## Model Architecture
 
-- Fine-tuned on iapp/thai_handwriting_dataset
-- Uses LoRA (r=8, alpha=16) targeting q_proj and v_proj attention layers
-- Trained for 3 epochs with batch size 4 and gradient accumulation 8
-- Learning rate 2e-4 with constant schedule and 3% warmup
+- Base: Llama-3.2-11B-Vision-Instruct
+- Adaptation: LoRA (r=8, alpha=16)
+- Target: q_proj, v_proj attention layers
+- Training: 3 epochs, batch size 4, gradient accumulation 8, lr 2e-4
 
-## Usage
+## Inference
 
 ```python
-from transformers import AutoModelForVision2Seq, AutoProcessor
+from transformers import AutoModelForVision2Seq, AutoProcessor, BitsAndBytesConfig
 from PIL import Image
 
 model_path = "Aekanun/thai-handwriting-llm"
+
+# BitsAndBytes config for efficient inference
+bnb_config = BitsAndBytesConfig(
+   load_in_4bit=True,
+   bnb_4bit_use_double_quant=True,
+   bnb_4bit_quant_type="nf4",
+   bnb_4bit_compute_dtype=torch.bfloat16
+)
+
+# Load processor and model
 processor = AutoProcessor.from_pretrained("meta-llama/Llama-3.2-11B-Vision-Instruct")
-model = AutoModelForVision2Seq.from_pretrained(model_path)
+model = AutoModelForVision2Seq.from_pretrained(
+   model_path,
+   device_map="auto",
+   torch_dtype=torch.bfloat16,
+   quantization_config=bnb_config
+)
 
 # Prepare input
 image = Image.open("handwriting.jpg")
@@ -49,7 +66,17 @@ messages = [
 ]
 
 # Generate text
-inputs = processor(text=processor.apply_chat_template(messages, tokenize=False), 
-                 images=image, return_tensors="pt")
-outputs = model.generate(**inputs, max_new_tokens=256)
+inputs = processor(
+   text=processor.apply_chat_template(messages, tokenize=False),
+   images=image, 
+   return_tensors="pt"
+)
+inputs = {k: v.to(model.device) for k, v in inputs.items()}
+
+outputs = model.generate(
+   **inputs,
+   max_new_tokens=256,
+   do_sample=False,
+   pad_token_id=processor.tokenizer.pad_token_id
+)
 text = processor.decode(outputs[0], skip_special_tokens=True)