Davis426/COMP8420-Healthcare-LLM-Assistant

Healthcare LLM Assistant - QLoRA fine-tunes

Two parallel QLoRA fine-tunes of small instruct models on the same 9,000-pair mix of public biomedical Q&A, served side-by-side in the parent project's Streamlit UI for a 3-way bake-off against GPT-5.5.

Variant	Subfolder	Base	Adapter	GGUF (Q4_K_M)
Qwen	qwen/	Qwen/Qwen2.5-1.5B-Instruct	qwen/qwen-medqa-adapter/ (~82 MB)	qwen/qwen-medqa-gguf/model.Q4_K_M.gguf (~941 MB)
Llama-3.2	llama32/	meta-llama/Llama-3.2-1B-Instruct	llama32/llama32-medqa-adapter/ (~50 MB)	llama32/llama32-medqa-gguf/model.Q4_K_M.gguf (~770 MB)

Both variants were trained with the same dataset, the same LoRA shape (r=16, α=32, all 7 projection layers) and the same SFT recipe, so any quality gap isolates the base-model effect.

Built as part of the COMP8420 (Macquarie University) main project on a healthcare NLP assistant. Companion code: https://github.com/NhatNguyen3001/COMP8420-Healthcare-LLM-Assistant (see the GitHub README for the full system: voice input, PII railguard, multi-agent RAG, evaluation notebooks.)

What is in this repo

.
├── qwen/
│   ├── qwen-medqa-adapter/                  # PEFT LoRA adapter
│   └── qwen-medqa-gguf/
│       ├── model.Q4_K_M.gguf                # Ollama-ready GGUF
│       └── Modelfile                        # Ollama registration recipe
└── llama32/
    ├── llama32-medqa-adapter/               # PEFT LoRA adapter
    └── llama32-medqa-gguf/
        ├── model.Q4_K_M.gguf
        └── Modelfile

The merged-but-unquantized safetensors is intentionally not uploaded for either variant; it is redundant for end users (use the GGUF for Ollama OR the adapter for transformers+peft).

Training data

9,000 question-answer pairs (train 8,100 / val 450 / test 450) drawn from six public sources, capped at 1,500 pairs per source for balance:

Source	Pairs	Notes
BioASQ (subset of training14b)	~1,500	factoid / list / summary biomedical Q&A
MedQuAD	~1,500	consumer-facing medical questions
DrugBank description	~1,500	"What is X?" templates
DrugBank indication	~1,500	indication / contraindication
DrugBank side_effects	~1,500	side-effect summaries
DrugBank mechanism_of_action	~1,500	MoA explanations

90 / 5 / 5 random split with seed=42. The OpenAI messages format is used at JSONL level; each variant's chat template (Qwen2.5 or Llama-3.1) is applied at training time, not stored in the JSONL.

Training setup

Same hyperparameters across both variants:

Hyperparameter	Value
LoRA rank r	16
LoRA alpha	32
LoRA target modules	all 7 projection layers (q, k, v, o, gate, up, down)
Max sequence length	1024
Per-device batch size	2
Gradient accumulation	4 (effective batch = 8)
Epochs	3
Learning rate	2e-4, cosine schedule
Optimizer	adamw_8bit
Seed	42
Hardware	RTX 4060 (8 GB, bf16)

Per-variant differences:

	Qwen	Llama-3.2
Base id	Qwen/Qwen2.5-1.5B-Instruct (4-bit NF4)	meta-llama/Llama-3.2-1B-Instruct (4-bit NF4)
Chat template	qwen-2.5	llama-3.1
Wall time (3 epochs)	~95 min	~30-45 min (smaller base)
Best val loss	1.5536 (~epoch 1.98)	see results/training_qlora_llama32.md in parent repo

Deployed checkpoints are end-of-epoch-3 for both (the "what a full QLoRA run gives you" baseline, not early-stopped).

Evaluation

Evaluated on the held-out 450-pair test set, with 100 stratified pairs (~17 per source) used as the common comparison sample across all evaluation notebooks.

Two evaluation passes:

1. Surface metrics: ROUGE-1/2/L + BERTScore-F1 (with the PubMedBERT backbone)
2. LLM-as-judge: GPT-5.4 scoring blind on Accuracy / Completeness / Clarity / Safety (0-10), reference-aware

Headline findings (vs GPT-5.5), Qwen variant:

• The Qwen QLoRA model wins ROUGE-L by +0.022 (+12% relative) and BERTScore-F1 by +0.0067 (+0.8% relative) against GPT-5.5
• The win is driven by template substitution, not factual improvement. The training set includes 71+ DrugBank entries sharing the skeleton "{X} pollen is the pollen of the {X} plant. {X} pollen is mainly used in allergenic testing." The fine-tune learns the template and slot-fills the entity at inference; ROUGE and BERTScore both reward this even when the substituted entity is wrong.
• Verified 0 / 450 literal Q+A pair overlap between train and test, so this is template generalization, not memorization.
• Under the LLM-as-judge Accuracy dimension, GPT-5.5 leads (judge results in the parent repo's results/llm_judge_evaluation.csv).

Llama-3.2 variant: see the 3-way numbers in the parent repo's results/model_comparison.csv (refreshed after the Llama run). The same template-substitution dynamic is expected on shared DrugBank slots; the contrast with the Qwen variant isolates the base-model contribution.

Detailed numbers and charts live in the parent repo:

• results/llm_generation_evaluation.csv + llm_generation_eval_chart.png + llm_generation_bertscore_chart.png
• results/llm_judge_evaluation.csv + llm_judge_eval_chart.png
• results/model_comparison.csv + model_comparison_chart.png
• results/qlora_loss_curve.png + results/qlora_source_mix.png

How to use

Replace <variant> with qwen or llama32 in the examples below.

Option 1: Ollama (recommended for local serving)

# Fetch one variant's GGUF + Modelfile
huggingface-cli download Davis426/COMP8420-Healthcare-LLM-Assistant \
  --include "qwen/qwen-medqa-gguf/*" \
  --local-dir ./models

# Register with Ollama
cd ./models/qwen/qwen-medqa-gguf
ollama create medqa-qwen -f Modelfile

# Try it
ollama run medqa-qwen "What is amoxicillin used for?"

For the Llama variant, swap every qwen for llama32 (paths) and the Ollama tag to medqa-llama32.

You can register both side-by-side; one ollama serve daemon handles both tags concurrently (OLLAMA_MAX_LOADED_MODELS defaults to 3).

Option 2: transformers + peft (Python)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# pick a variant
base_id    = "Qwen/Qwen2.5-1.5B-Instruct"
subfolder  = "qwen/qwen-medqa-adapter"
# or:
# base_id   = "meta-llama/Llama-3.2-1B-Instruct"
# subfolder = "llama32/llama32-medqa-adapter"
adapter_id = "Davis426/COMP8420-Healthcare-LLM-Assistant"

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id, subfolder=subfolder)

messages = [{"role": "user", "content": "What is amoxicillin used for?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Option 3: llama.cpp directly

huggingface-cli download Davis426/COMP8420-Healthcare-LLM-Assistant \
  --include "qwen/qwen-medqa-gguf/model.Q4_K_M.gguf" --local-dir .

./llama-cli -m qwen/qwen-medqa-gguf/model.Q4_K_M.gguf \
  -p "What is amoxicillin used for?" -n 256

Limitations

Both models are teaching / research artifacts. Do not use for real clinical decisions. Specifically:

• Catastrophic forgetting on out-of-distribution prompts. Fine-tuning on a narrow Q&A distribution at the 1-1.5B parameter scale shifts each base model hard. Casual / non-medical questions get answered in MedQA-style; the base model's general conversational ability is degraded.
• Weakened in-context grounding. Every training pair has shape user_question -> answer, with no retrieved-context block. As a result both fine-tuned models partly lose the ability to read RAG passages in the prompt and tend to answer from parametric memory even when correct evidence is supplied. The parent repo's MASS-RAG pipeline retains GPT-5.5 for cases where grounded answers matter; the local models are sidebar-selectable for the comparison experience.
• No factual safety net. Both training data and evaluation rely on existing biomedical corpora; the models have no live knowledge cutoff or up-to-date drug-interaction database. The parent repo applies a regex-based PII railguard on user input, but model output itself is not safety-filtered beyond what each base model already does.
• English only.
• Llama-3.2 base licence: Llama-3.2 community licence applies to the Llama variant (acceptance via the gated HF repo); see the Meta licence for permitted uses.

License

The fine-tuned adapters and GGUFs in this repo are released under cc-by-nc-4.0 (research and non-commercial use). Base model licences override where stricter: Qwen2.5 is Apache-2.0; Llama-3.2 is under the Meta Llama 3.2 Community Licence. Downstream dataset licences may impose additional restrictions; please consult each source (BioASQ, MedQuAD, DrugBank, MedRAG textbooks) before redistribution.

Citation

If you use or build on this work, please reference:

@misc{comp8420-2026-medqa,
  title  = {Healthcare NLP Assistant: parallel QLoRA fine-tunes of Qwen2.5-1.5B and Llama-3.2-1B for medical Q&A},
  author = {Davis426},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/Davis426/COMP8420-Healthcare-LLM-Assistant}}
}

Built on top of:

• Qwen2.5 (Alibaba): https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct
• Llama-3.2 (Meta): https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct
• QLoRA (Dettmers et al., 2023): https://arxiv.org/abs/2305.14314
• MASS-RAG (Xiao, Huang, Liu, Xie, 2026): https://arxiv.org/abs/2604.18509 (used by the parent repo's retrieval pipeline that these models plug into)
• Unsloth: https://github.com/unslothai/unsloth
• llama.cpp + Ollama for GGUF serving