# Model Card for BioGPT-FineTuned-MedicalTextbooks-FP16

# Model Overview
This model is a fine-tuned and quantized version of the microsoft/biogpt model, specifically tailored for medical text understanding. It was fine-tuned on the dmedhi/medical-textbooks dataset from Hugging Face and subsequently quantized to FP16 (half-precision) to reduce memory usage and improve inference speed while maintaining accuracy. The model is designed for tasks like keyword extraction from medical texts and generative tasks in the biomedical domain.

# Model Details
```
Base Model: microsoft/biogpt
Fine-Tuning Dataset: dmedhi/medical-textbooks (15,970 rows)
Quantization: FP16 (half-precision) using PyTorch's .half() method
Model Type: Causal Language Model
Language: English
```
# Intended Use
This model is intended for:

- Keyword Extraction: Extracting relevant lines containing specific keywords (e.g., "anatomy") from medical textbooks, along with metadata like book names.
- Generative Tasks: Generating short explanations or summaries in the biomedical domain (e.g., answering questions like "What is anatomy?").
- Research and Education: Assisting researchers, students, and educators in exploring medical texts and generating insights.
# Out of Scope
- Real-time clinical decision-making or medical diagnosis (not evaluated for such tasks).
- Non-English text processing (not tested on other languages).
- Tasks requiring high precision in generative output without human oversight.
# Training Details
# Dataset
The model was fine-tuned on the dmedhi/medical-textbooks dataset, which contains excerpts from medical textbooks with two attributes:

**text:** The content of the excerpt.
**book:** The name of the book (e.g., "Gray's Anatomy").
# Dataset Splits:
- Original split: train (15,970 rows).
- Custom splits: 80% train (12,776 rows), 20% validation (3,194 rows).
# Training Procedure
# Preprocessing:

- Tokenized the text field using the BioGPT tokenizer (microsoft/biogpt).
- Set max_length=512, with truncation and padding.
- Used input_ids as labels for causal language modeling.
# Fine-Tuning:
- Fine-tuned microsoft/biogpt using Hugging Face's Trainer API.
```
Training arguments:
Epochs: 1
Batch size: 4 per device
Learning rate: 2e-5
Mixed precision: FP16 (fp16=True)
Evaluation strategy: Steps (every 1000 steps)
Training loss decreased from 2.8409 to 2.7006 over 3,194 steps.
Validation loss decreased from 2.7317 to 2.6512.
```
# Quantization:
- Converted the fine-tuned model to FP16 using PyTorch's .half() method.
- Saved as ./biogpt_finetuned/final_model_fp16.
- Compute Infrastructure
- Hardware: 12 GB GPU (NVIDIA)
- Environment: Jupyter Notebook on Windows
- Framework: PyTorch, Hugging Face Transformers
- Training Time: Approximately 27 minutes for 1 epoch
# Evaluation
**Metrics**
```
Training Loss: Decreased from 2.8409 to 2.7006.
Validation Loss: Decreased from 2.7317 to 2.6512.
Memory Usage: Post-quantization memory usage reported as ~661 MB (FP16), though actual savings may vary due to buffers and non-weight tensors.
```
# Qualitative Testing
**Generative Task:** Generated a response to "What is anatomy?" with reasonable output: "What is anatomy? Anatomy is the basis of medicine..."
**Keyword Extraction:** Successfully extracted up to 10 lines containing keywords (e.g., "anatomy") with corresponding book names (e.g., "Gray's Anatomy").

# Usage
**Installation**
- Ensure you have the required libraries installed:

```
pip install transformers torch datasets sacremoses
```
# Loading the Model
- Load the quantized FP16 model and tokenizer:
```
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "path/to/biogpt_finetuned/final_model_fp16"  # Update with your HF repo path
model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()
```
# Example 1: Generative Inference
# Generate text with the quantized model:
```

input_text = "What is anatomy?"
inputs = tokenizer(input_text, return_tensors="pt", padding=True, truncation=True, max_length=512)
inputs = {k: v.to(device) for k, v in inputs.items()}
with torch.no_grad():
    outputs = model.generate(**inputs, max_length=50)
output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(output_text)
```
# Example 2: Keyword Extraction
```
from datasets import load_from_disk

original_datasets = load_from_disk('path/to/original_medical_textbooks')

def extract_lines_with_keyword(keyword, dataset_split='train', max_results=10):
    dataset = original_datasets[dataset_split]
    matching_lines = []
    for entry in dataset:
        text = entry['text']
        book = entry['book']
        lines = text.split('\n')
        for line in lines:
            if keyword.lower() in line.lower():
                matching_lines.append({'text': line.strip(), 'book': book})
                if len(matching_lines) >= max_results:
                    return matching_lines
    return matching_lines

keyword = "anatomy"
matching_lines = extract_lines_with_keyword(keyword)
for i, match in enumerate(matching_lines, 1):
    print(f"{i}. Text: {match['text']}")
    print(f"   Book: {match['book']}\n")
```
# Limitations
- Quantization Trade-offs: FP16 quantization may lead to minor accuracy degradation, though not extensively evaluated.
- Dataset Bias: Fine-tuned only on dmedhi/medical-textbooks, which may not cover all medical domains or topics.
- Generative Quality: Generative outputs may require human oversight for correctness.
- Scalability: Keyword extraction relies on string matching, not semantic understanding, limiting its ability to capture nuanced relationships.