Bioinformatics LLM Fine-Tuning with LoRA

This repository demonstrates fine-tuning a lightweight LLM (TinyLlama-1.1B) on bioinformatics instruction-response prompts using Low-Rank Adaptation (LoRA). The goal is to adapt a general instruction-tuned model to answer domain-specific questions and showcase the improvements before and after fine-tuning.

---

Objective

• Fine-tune TinyLlama for bioinformatics instructions
• Run locally on Apple Silicon (M4 Max, MPS backend)
• Compare outputs before and after LoRA fine-tuning

---

Repository Structure

bioinfo-lora-finetuning-demo/
├── data/
│   └── bioinfo_train.jsonl          # Instruction-response dataset
├── src/
│   ├── lora_train.py                # LoRA fine-tuning script
│   ├── lora_infer_before.py         # Baseline inference
│   ├── lora_infer_after.py          # Inference using LoRA adapter
│   └── merge_lora.py                # Merge adapter into base model
├── results/
│   ├── sample_outputs_before.txt
│   ├── sample_outputs_after.txt
│   └── merged-model/
├── requirements.txt
└── README.md

---

Environment Setup

1. Clone the repository and create a virtual environment:

git clone <your-repo-url>
cd bioinfo-lora-finetuning-demo
python -m venv venv
source venv/bin/activate

2. Install dependencies:

pip install --upgrade pip setuptools wheel
pip install -r requirements.txt

Note: On macOS M4, if you encounter sentencepiece build errors, use:

pip install sentencepiece --prefer-binary

---

Dataset

• data/bioinfo_train.jsonl contains bioinformatics instruction-response pairs in JSON Lines format:

{"instruction": "Explain what a FASTQ file is.", "output": "A FASTQ file stores sequencing reads with quality scores..."}
{"instruction": "What is SNP annotation?", "output": "SNP annotation links single-nucleotide polymorphisms to genes and predicts functional impacts."}

• You can expand this dataset to hundreds of examples for improved results.

---

Fine-Tuning

python src/lora_train.py \     
  --dataset_path data/bioinfo_train.jsonl \
  --epochs 3 \
  --batch_size 2 \
  --gradient_accumulation 4 \
  --lr 2e-4 \
  --output_dir results/lora-adapter

• Saves LoRA adapter to ./results/lora-adapter
• Uses MPS GPU if available
• Training logs show decreasing loss over steps

---

Inference

Baseline (Before Fine-Tuning)

python src/lora_infer_before.py

Fine-Tuned Model (After LoRA)

python src/lora_infer_after.py

Example comparison:

Prompt	Baseline	LoRA Fine-Tuned
Explain FASTQ	“A text file containing sequences.”	“A FASTQ file stores sequencing reads with quality scores, used in genome sequencing and bioinformatics pipelines.”
SNP annotation	“A process in genomics.”	“SNP annotation links single-nucleotide polymorphisms to genes and predicts functional impact.”

---

Merge LoRA Adapter

To create a standalone fine-tuned model:

python src/merge_lora.py

• Output: ./results/merged-model/
• Load without PEFT adapters:

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("./results/merged-model").to("mps")
tokenizer = AutoTokenizer.from_pretrained("./results/merged-model")

---

Results

• Train time: ~1 min per epoch on M4 Max
• Loss decrease: ~10 → 0.3
• Output improvement: Domain-specific answers with bioinformatics terms