amornpan/openthaigpt1.5-14b-MedChatModelV1

✨ Fine-tuning the MedChat model for GPU efficiency. ✨

🇹🇭 Model Card for openthaigpt1.5-14b-medical-tuned

ℹ️ This version is optimized for GPU. Please wait for the CPU version, which will be available soon.!!

This model is fine-tuned from openthaigpt1.5-14b-instruct using Supervised Fine-Tuning (SFT) on the Thaweewat/thai-med-pack dataset. The model is designed for medical question-answering tasks in Thai, specializing in providing accurate and contextual answers based on medical information.

Model Description

This model was fine-tuned using Supervised Fine-Tuning (SFT) to optimize it for medical question answering in Thai. The base model is openthaigpt1.5-14b-instruct, and it has been enhanced with domain-specific knowledge using the Thaweewat/thai-med-pack dataset.

• Model type: Causal Language Model (AutoModelForCausalLM)
• Language(s): Thai
• License: Apache License 2.0
• Fine-tuned from model: openthaigpt1.5-14b-instruct
• Dataset used for fine-tuning: Thaweewat/thai-med-pack

Model Sources

• Repository: https://huggingface.co/amornpan
• Citing Repository: https://huggingface.co/Aekanun
• Base Model: https://huggingface.co/openthaigpt/openthaigpt1.5-14b-instruct
• Dataset: https://huggingface.co/datasets/Thaweewat/thai-med-pack

Uses

Direct Use

The model can be directly used for generating medical responses in Thai. It has been optimized for:

• Medical question-answering
• Providing clinical information
• Health-related dialogue generation

Downstream Use

This model can be used as a foundational model for medical assistance systems, chatbots, and applications related to healthcare, specifically in the Thai language.

Out-of-Scope Use

• This model should not be used for real-time diagnosis or emergency medical scenarios.
• Avoid using it for critical clinical decisions without human oversight, as the model is not intended to replace professional medical advice.

Bias, Risks, and Limitations

Bias

• The model might reflect biases present in the dataset, particularly when addressing underrepresented medical conditions or topics.

Risks

• Responses may contain inaccuracies due to the inherent limitations of the model and the dataset used for fine-tuning.
• This model should not be used as the sole source of medical advice.

Limitations

• Limited to the medical domain.
• The model is sensitive to prompts and may generate off-topic responses for non-medical queries.

Model Training Results

985/985 8:34:43, Epoch 134/141]
Step	Training Loss	Validation Loss
50	1.883700	1.708532
100	1.792500	1.528184
150	1.555000	1.296583
200	1.403900	1.251281
250	1.374300	1.225630
300	1.321000	1.195238
350	1.313900	1.187670
400	1.299000	1.181292
450	1.296400	1.177670
500	1.285000	1.173616
550	1.272800	1.170705
600	1.251200	1.169226
650	1.262600	1.166078
700	1.255300	1.165633
750	1.251600	1.165041
800	1.252300	1.162943
850	1.232700	1.164691
900	1.247300	1.163449
950	1.246300	1.163610

How to Get Started with the Model

Here’s how to load and use the model for generating medical responses in Thai:

Using Google Colab Pro or Pro+ for fine-tuning and inference.

1. Install the Required Packages

First, ensure you have installed the required libraries by running:

! pip install --upgrade torch transformers accelerate bitsandbytes --upgrade

2. Load the Model and Tokenizer

You can load the model and tokenizer directly from Hugging Face using the following code:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

Define the model path

model_path = 'amornpan/openthaigpt1.5-14b-MedChatModelV1'

Load the tokenizer and model

tokenizer = AutoTokenizer.from_pretrained(model_path)
tokenizer.pad_token = tokenizer.eos_token

3. Prepare Your Input (Custom Prompt)

Create a custom medical prompt that you want the model to respond to:

custom_prompt = "โปรดอธิบายลักษณะช่องปากที่เป็นมะเร็งในระยะเริ่มต้น"
PROMPT = f'[INST] <You are a question answering assistant. Answer the question as truthfully and helpfully as possible. คุณคือผู้ช่วยตอบคำถาม จงตอบคำถามอย่างถูกต้องและมีประโยชน์ที่สุด<>{custom_prompt}[/INST]'

# Tokenize the input prompt
inputs = tokenizer(PROMPT, return_tensors="pt", padding=True, truncation=True)

4. Configure the Model for Efficient Loading (4-bit Quantization)

The model uses 4-bit precision for efficient inference. Here’s how to set up the configuration:

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

5. Load the Model with Quantization Support

Now, load the model with the 4-bit quantization settings:

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    quantization_config=bnb_config,
    trust_remote_code=True
)

6. Move the Model and Inputs to the GPU (prefer GPU)

For faster inference, move the model and input tensors to a GPU, if available:

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
inputs = {k: v.to(device) for k, v in inputs.items()}

7. Generate a Response from the Model

Now, generate the medical response by running the model:

outputs = model.generate(**inputs, max_new_tokens=256, do_sample=True)

8. Decode the Generated Text

Finally, decode and print the response from the model:

generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)

9. Output

[INST] <You are a question answering assistant. Answer the question as truthfully and helpfully as possible.
คุณคือผู้ช่วยตอบคำถาม จงตอบคำถามอย่างถูกต้องและมีประโยชน์ที่สุด<>
โปรดอธิบายลักษณะช่องปากที่เป็นมะเร็งในระยะเริ่มต้น[/INST]

สำหรับมะเร็งช่องปากในระยะแรกอาจรวมถึงอาการหรือลักษณะดังต่อไปนี้:
1. แผลบนช่องปากหรือกระพูดนิ่มที่อยู่กับที่และไม่หายไปแม้จะผ่านการรักษาด้วยตนเอง
2. บวมที่อยู่กับที่ที่ข้างใดข้างหนึ่งของริมฝีปาก
3. แผลเปื่อยหรือพังผืดที่เกิดขึ้นที่กระพูดหรือฟันที่ไม่หาย
4. ความเปลี่ยนแปลงของผิว pigment ในช่องปาก เช่น สีของกระพืดหรือริมฝีปากที่เปลี่ยนเป็นสีขาวหรือขาว
5. ปัญหาในการพูดหรือกินอาหาร
6. ขดลวดที่ด้านข้างหรือใต้คอที่เจริญเติบโต
7. อาการเจ็บจี๊ด

👤 Authors

• Amornpan Phornchaicharoen ([email protected])
• Aekanun Thongtae ([email protected])
• Montita Somsoo ([email protected]
• Phongsatorn Somjai ([email protected])
• Jiranuwat Songpad ([email protected])
• Benchawan Wangphoomyai ([email protected])