Medical-llama-2

Medical Chatbot Model Card

Model Details

Model Description

This is a fine-tuned version of the LLaMA 2-based language model for medical chatbot purposes. The model has been trained on a medical dataset to provide information about various medical terms, conditions, treatments, and diseases in response to user queries.

It has been fine-tuned using Supervised Fine-Tuning (SFT) with the Lora (Low-Rank Adaption) method to optimize for medical conversational tasks. It is capable of providing accurate and relevant responses for medical queries.

• Developed by: Abhinandan Patil
• License: [Specify License Type, e.g., "MIT" or "Apache 2.0"]
• Model type: Causal Language Model (Fine-tuned LLaMA 2)
• Language(s): English
• Finetuned from: Base model: "LLaMA 2"
• Training Task: Medical Question Answering, Medical Chatbot

Model Sources

• Repository: https://huggingface.co/patikop/medical-chatbot
• Demo: [Link to live demo or API endpoint, if available]

Uses

Direct Use

This model can be used directly as a conversational agent for medical-related queries. By feeding it a user prompt, it generates informative and contextually relevant responses regarding medical conditions, symptoms, treatments, etc.

Downstream Use

This model can also be used in other applications where medical information is needed, such as:

• Medical virtual assistants
• Medical FAQ chatbots
• Healthcare support applications

Out-of-Scope Use

This model should NOT be used for:

• Medical Diagnosis: The model is not intended to replace healthcare professionals.
• Critical Medical Decision-Making: It should not be used in scenarios where human judgment is essential (e.g., emergency medicine or surgery).
• Legal Advice: The model is not trained to provide any legal advice related to healthcare or medical practices.

Bias, Risks, and Limitations

Risks

• Misinformation: While the model has been fine-tuned on medical text, there is a possibility of generating incorrect or outdated medical information. The model should always be used with proper caution, and outputs should be validated by qualified medical professionals.
• Biases: The model's responses are influenced by the dataset it was trained on. If the training data had inherent biases, the model may propagate those biases in its responses.
• Ethical Concerns: Care should be taken when using this model in applications that interact with vulnerable populations, as inappropriate or incorrect information could have serious consequences.

Recommendations

• Validation: Always validate medical responses with expert healthcare professionals before making decisions based on the model's output.
• Transparency: Users should be informed that the model is not a replacement for medical advice from qualified professionals.
• Monitor Updates: Continuously monitor the model’s outputs and retrain it on updated datasets to keep it current.

Training Details

Training Data

The model was fine-tuned using a custom dataset of medical terms, which includes information about diseases, symptoms, treatments, and medical terminology. The dataset was formatted for use in Supervised Fine-Tuning (SFT) tasks, ensuring the model learned to generate relevant medical responses based on user inputs.

Training Procedure

• Preprocessing: The data was cleaned and tokenized using the LLaMA tokenizer. Padding was handled with respect to the maximum sequence length of 512 tokens to ensure consistency in input length for training.

Hyperparameters:

• Batch size: 2
• Steps: 100
• Optimizer: AdamW
• TrainOutput: (global_step=100, training_loss=1.62529541015625, metrics={'train_runtime': 907.0293, 'train_samples_per_second': 0.221, 'train_steps_per_second': 0.11, 'total_flos': 3848267600855040.0, 'train_loss': 1.62529541015625, 'epoch': 0.02914602156805596})

---

Evaluation

Testing Data, Factors & Metrics

The model was tested using held-out medical texts and queries to ensure that it provides relevant and coherent answers. Key evaluation metrics include:

• Accuracy: Measures the relevance and correctness of the answers generated by the model.
• Perplexity: Measures the fluency and coherence of the text generated by the model, assessing how well the model predicts the next word in a sequence.

Results

The model performs well for general medical queries, offering coherent and accurate answers. However, for complex or rare medical cases, the model's responses should be validated by medical professionals.

---

Model Examination

The model’s responses should be carefully examined for bias and factual correctness before being deployed in any clinical or high-stakes environment. Additionally, performance should be analyzed across different medical domains (e.g., oncology, cardiology) to ensure that the model provides consistent and reliable results across various fields.

---

Environmental Impact

Training large language models involves significant computational resources, which can result in substantial environmental impact. You can estimate your model’s carbon footprint using the Machine Learning Impact Calculator to better understand the emissions associated with the training process.

---

Technical Specifications

Model Architecture

The model is based on the LLaMA 2 architecture, a causal language model, and was fine-tuned using Supervised Fine-Tuning (SFT) with Low-Rank Adaption (Lora).

How to Get Started with the Model

You can quickly start using the model via the Hugging Face library. Here’s how to load and interact with the model:

Installation Requirements

altair==4.2.2
streamlit==1.19.0
transformers==4.33.2
torch==2.0.0
accelerate==0.20.3

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

# Load the model and tokenizer from Hugging Face
hf_model = AutoModelForCausalLM.from_pretrained("patikop/medical-chatbot")
hf_tokenizer = AutoTokenizer.from_pretrained("patikop/medical-chatbot")

# Create the pipeline
hf_pipeline = pipeline(task="text-generation", model=hf_model, tokenizer=hf_tokenizer, max_length=300)

# Define your query
user_prompt = "What is malaria?"

# Generate the response
response = hf_pipeline(f"<s>[INST] {user_prompt} [/INST]")

# Print the response
print(response[0]['generated_text'])