alzbot-qwen3-4b-lora

AlzBot: Alzheimer's Disease Research Assistant (LoRA Adapter)

A domain-specific LoRA adapter fine-tuned on 95,374 biomedical records (~110M tokens) from 13 public data sources, specializing in Alzheimer's disease research.

Model Details

Data Sources

The AlzCorpus was constructed from 13 public biomedical APIs:

• Europe PMC (75,954 articles) — research abstracts and open-access full text
• openFDA (25,100) — drug labels, indications, warnings
• PubMed (9,994) — articles with metadata
• ClinicalTrials.gov (4,194) — clinical trial records
• SIDER (1,226) — drug side effects
• Open Targets (868) — drug-target associations
• Plus: ChEMBL, PharmGKB, OMIM, and 6,104 conversational Q&A pairs

All records are evidence-graded (A-E) based on study type, sample size, and recency.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-4B-Instruct-2507",
    dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
model = PeftModel.from_pretrained(base, "pedrobsb/alzbot-qwen3-4b-lora")
model.eval()

messages = [
    {"role": "system", "content": "You are AlzBot, a specialized Alzheimer's disease research assistant."},
    {"role": "user", "content": "What are the FDA-approved treatments for Alzheimer's disease?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=512, temperature=0.3, top_p=0.9, do_sample=True)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Training Configuration

• LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Optimizer: Paged AdamW 8-bit
• Learning rate: 1e-4 (cosine schedule, 500 warmup steps)
• Batch size: 16 (effective)
• Sequence length: 512 tokens (with packing)
• Precision: float16
• Epochs: 2
• GPU: NVIDIA A10G (24 GB VRAM), AWS g5.2xlarge

Eval Results

Topics Covered

• FDA-approved treatments (donepezil, rivastigmine, galantamine, memantine, lecanemab, donanemab)
• Genetic risk factors (APOE e4, PSEN1, APP)
• Biomarkers (amyloid-beta, tau, CSF markers, PET imaging)
• Clinical trials and emerging therapies
• Drug interactions and contraindications
• Caregiver support and patient care
• Evidence grading and source citations

Citation

@article{albuquerque2026alzbot,
  title={AlzBot: Toward Community-Driven Lightweight Domain Language Models as Living Alternatives to Systematic Literature Reviews},
  author={Albuquerque, Pedro H. M.},
  year={2026}
}

Disclaimer

This model is designed as a research tool for clinicians and researchers, not as a direct-to-patient advisory system. It should not be used for individual treatment decisions without verification by a qualified clinician.

Framework Versions

• PEFT 0.19.1
• Transformers 5.12.1
• PyTorch 2.12.1+cu130
• TRL 1.7.0