Skill: Low-Resource AI Researcher

ID: 215
Category: AI/ML Research
Language: Python
Framework: PyTorch + PEFT (LoRA/QLoRA) + Transformers

Overview

Based on Parameter-Efficient Fine-Tuning (PEFT) technology, trains high-performance medical domain large language models on consumer-grade GPUs or single A100. Supports advanced fine-tuning methods such as LoRA, QLoRA, optimized for medical text understanding and generation tasks.

Features

• 🚀 Parameter-Efficient Fine-Tuning: LoRA, QLoRA, DoRA support
• 🏥 Medical Domain Optimized: Pre-configured for medical QA, diagnosis, clinical notes
• 💻 Low-Resource Ready: Optimized for consumer GPUs (RTX 3090/4090) and single A100
• 📊 Quantization: 4-bit/8-bit quantization with bitsandbytes
• 🔄 Multi-Task: Supports SFT, DPO, and medical instruction tuning
• 📝 Medical Datasets: Built-in support for PubMedQA, MedQA, MIMIC-III

Installation

# Core dependencies
pip install torch transformers datasets accelerate peft bitsandbytes

# Optional for training optimization
pip install flash-attn --no-build-isolation
pip install wandb tensorboard

# Medical NLP utilities
pip install scispacy scikit-learn

Quick Start

from skills.low_resource_ai_researcher.scripts.main import MedicalPEFTTrainer

# Initialize trainer
trainer = MedicalPEFTTrainer(
    model_name="meta-llama/Llama-2-7b-hf",
    task="medical_qa"
)

# Train with LoRA
trainer.train(
    output_dir="./medical_lora_model",
    num_epochs=3,
    batch_size=4,
    use_qlora=True  # 4-bit quantization
)

Configuration

Hardware Profiles

LoRA Config

lora:
  r: 64              # LoRA rank
  lora_alpha: 128    # Scaling factor
  target_modules:    # Modules to apply LoRA
    - q_proj
    - v_proj
    - k_proj
    - o_proj
    - gate_proj
    - up_proj
    - down_proj
  lora_dropout: 0.05
  bias: "none"
  task_type: "CAUSAL_LM"

CLI Usage

# Basic training
python scripts/main.py \
    --model_name_or_path meta-llama/Llama-2-7b-hf \
    --dataset medical_qa \
    --output_dir ./output \
    --use_qlora \
    --per_device_train_batch_size 4

# With custom config
python scripts/main.py --config configs/medical_qlora.yaml

# Resume training
python scripts/main.py --resume_from_checkpoint ./output/checkpoint-1000

API Reference

MedicalPEFTTrainer

trainer = MedicalPEFTTrainer(
    model_name: str,              # Base model name/path
    task: str,                    # Task type: medical_qa, diagnosis, clinical_note
    lora_r: int = 64,             # LoRA rank
    lora_alpha: int = 128,        # LoRA alpha
    use_qlora: bool = False,      # Use 4-bit quantization
    target_modules: List[str] = None,
    device_map: str = "auto",
    trust_remote_code: bool = True
)

Methods

Supported Models

• LLaMA 2/3 (7B, 13B, 70B)
• Mistral (7B, 8x7B)
• Yi (6B, 34B)
• Qwen (7B, 14B, 72B)
• Baichuan (7B, 13B)
• ChatGLM (6B)

Medical Datasets

Performance Benchmarks

Best Practices

1. Gradient Accumulation: Use for effective larger batch sizes
2. Learning Rate: Start with 2e-4 for LoRA, 1e-4 for full fine-tuning
3. Warmup Steps: 100 steps for medical domain adaptation
4. Max Length: 2048-4096 for clinical notes, 512-1024 for QA
5. Data Quality: Filter out low-quality medical data carefully

Troubleshooting

Out of Memory

# Enable gradient checkpointing
trainer.train(gradient_checkpointing=True)

# Reduce sequence length
trainer.train(max_seq_length=1024)

# Use DeepSpeed ZeRO-3 for large models

Slow Training

# Enable Flash Attention
trainer.train(use_flash_attention=True)

# Use bf16 on Ampere GPUs
trainer.train(bf16=True)

License

This skill follows the license of the underlying models used. Medical applications require compliance with HIPAA/GDPR regulations.

References

1. Hu et al. (2021) - LoRA: Low-Rank Adaptation of Large Language Models
2. Dettmers et al. (2023) - QLoRA: Efficient Finetuning of Quantized LLMs
3. Singhal et al. (2023) - Large Language Models Encode Clinical Knowledge

Risk Assessment

Security Checklist

• No hardcoded credentials or API keys
• No unauthorized file system access (../)
• Output does not expose sensitive information
• Prompt injection protections in place
• Input file paths validated (no ../ traversal)
• Output directory restricted to workspace
• Script execution in sandboxed environment
• Error messages sanitized (no stack traces exposed)
• Dependencies audited

Prerequisites

# Python dependencies
pip install -r requirements.txt

Evaluation Criteria

Success Metrics

• Successfully executes main functionality
• Output meets quality standards
• Handles edge cases gracefully
• Performance is acceptable

Test Cases

1. Basic Functionality: Standard input → Expected output
2. Edge Case: Invalid input → Graceful error handling
3. Performance: Large dataset → Acceptable processing time

Lifecycle Status

• Current Stage: Draft
• Next Review Date: 2026-03-06
• Known Issues: None
• Planned Improvements:
  • Performance optimization
  • Additional feature support