Why Fine-Tune a Model?

Large language models like Mistral 7B are trained on broad, internet-scale data. That makes them capable generalists — but generalists have limitations when applied to specialized domains.

Fine-tuning is the process of continuing the training of a pre-trained model on a smaller, domain-specific dataset. Instead of learning from scratch (which requires massive compute and data), we teach an already-capable model to speak the language of a specific field — in this case, breast cancer medicine.

The goal is not to make the model "smarter" in a general sense, but more:

• Accurate — using the right clinical terminology and referencing real medical concepts
• Consistent — answering breast cancer questions in the structured, informative style of medical Q&A
• Relevant — focusing its generation on domain knowledge rather than generic preambles

Why QLoRA?

Training all 7 billion parameters requires dozens of GB of VRAM and days of compute — out of reach without expensive hardware. QLoRA (Quantized Low-Rank Adaptation) makes it accessible with two techniques:

1. 4-bit quantization — model weights are compressed from 32-bit floats to 4-bit integers, reducing VRAM from ~28 GB to ~5 GB with minimal quality loss.
2. LoRA adapters — instead of updating all 7B parameters, small trainable matrices (adapters) are injected into the attention layers. Only ~1–2% of total parameters are trained. The rest stay frozen.

The result: a meaningful domain fine-tune on a single T4 GPU (Google Colab free tier) in approximately 35 minutes.

---

Model Details

---

Training Data

Trained on DiegoDomLarr/breast-cancer-qa — 1,061 breast cancer Q&A pairs from two sources:

Filter keywords: breast cancer, breast carcinoma, BRCA1, BRCA2, HER2, tamoxifen, mastectomy, lumpectomy, mammogram, ductal carcinoma, lobular carcinoma, triple negative breast, aromatase inhibitor, trastuzumab

Dataset statistics:

• Average question length: 541 characters
• Average answer length: 621 characters

Training format — all examples were converted to Mistral's instruction template:

markdown
Copy code
<s>[INST] {question} [/INST] {answer} </s>

---

Evaluation

The fine-tuned model was compared against the base Mistral-7B-Instruct on the prompt:

"What are the main risk factors for breast cancer?"

Fine-tuned model response:

Several factors can increase the risk of developing breast cancer. Here are some of the most common risk factors:

1. Gender: Being female is the greatest risk factor for breast cancer.
2. Age: The risk of breast cancer increases as women get older. Most breast cancers are diagnosed in women over the age of 50.
3. Genetic Factors: Certain genetic mutations, such as those in the BRCA1 and BRCA2 genes, can significantly increase the risk. Women with a family history of breast cancer in first-degree relatives are also at higher risk.
4. Lifestyle Factors: A lack of physical activity, a diet high in saturated fat, being overweight or obese, and smoking all contribute to increased risk.

Base Mistral response:

Breast cancer is the most common cancer among women worldwide. Several risk factors can increase a woman's chance of developing breast cancer.

1. Age: The risk increases as women get older. Most breast cancers are diagnosed after age 50.
2. Genetic factors: A family history of breast cancer increases the risk. Inherited mutations in BRCA1 and BRCA2 significantly increase risk.
3. Hormonal factors: Extended exposure to estrogen and progesterone can increase risk. Factors include early menstruation, late menopause, and never having given birth.

ROUGE-L score (base vs fine-tuned): 0.4509

A ROUGE-L of ~0.45 confirms the fine-tuned model generates answers that are meaningfully different from the base — more structured, more patient-oriented, and covering additional factors (gender, lifestyle) not prominently addressed by the base model.

---

How to Use

Requirements

bash
Copy code
pip install transformers peft bitsandbytes accelerate

Inference

python
Copy code
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

base_model_id = "mistralai/Mistral-7B-Instruct-v0.2"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

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

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, "DiegoDomLarr/mistral-7b-breast-cancer-qlora")
model.eval()
model.config.use_cache = True

prompt = "<s>[INST] What are the side effects of tamoxifen? [/INST]"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=300,
        temperature=0.7,
        do_sample=True,
    )

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

---

Limitations and Intended Use

This model is for educational and research purposes only. It is not a medical device and must not be used for clinical diagnosis, treatment decisions, or patient care.

• Responses may contain inaccuracies or outdated medical information — always verify with a licensed healthcare professional.
• The model was trained on ~1,000 examples, which is small by fine-tuning standards. It may hallucinate or generalize poorly on edge-case questions.
• Coverage is limited to breast cancer topics represented in the training data.
• This model has not been audited, validated, or certified for any medical use.

---

Tech Stack

---

About This Project

This model was built as an end-to-end learning project covering the full fine-tuning pipeline:

1. Curating and publishing a domain-specific dataset to HF Hub
2. Loading a 7B model in 4-bit on consumer hardware (Colab T4)
3. Applying LoRA adapters with PEFT
4. Training with SFTTrainer (TRL library)
5. Evaluating with ROUGE-L against the base model
6. Publishing the adapter and model card to HuggingFace Hub

Author: DiegoDomLarr Dataset: DiegoDomLarr/breast-cancer-qa Base model: mistralai/Mistral-7B-Instruct-v0.2

---

License

Apache 2.0 — same as the base model.