trialmatch-gemma4

trialmatch-gemma4

LoRA fine-tune of Gemma 4 E4B for clinical trial eligibility pre-screening. Given a patient profile and a trial's eligibility criteria, the model returns a structured 5-field verdict used inside TrialMatch — a fully offline, CPU-only clinical trial matching app.

Model Details

Model Description

trialmatch-gemma4 is a LoRA adapter trained on top of unsloth/gemma-4-E4B-it-unsloth-bnb-4bit. It is fine-tuned for a single task: reading a structured patient profile + clinical reasoning summary + trial eligibility criteria text and returning a structured eligibility verdict.

The model outputs exactly five fields:

VERDICT: MATCH | PARTIAL | NO
CONFIDENCE: 0–100
REASON: one plain-English sentence
DISQUALIFIERS: exact criterion text, or NONE
NEXT STEP: what the patient should do next

• Developed by: Bowei (@boweiiismyname)
• Model type: Causal LM — LoRA adapter (text-only fine-tune on a multimodal base)
• Language(s): English
• License: Apache 2.0
• Finetuned from: unsloth/gemma-4-E4B-it-unsloth-bnb-4bit

Model Sources

• Repository: Boweii22/TrialMatch
• Demo: TrialMatch runs fully offline — see the repo for setup instructions

---

Uses

Direct Use

Load the LoRA adapter on top of Gemma 4 E4B and pass a patient profile + eligibility criteria to get a structured MATCH/PARTIAL/NO verdict. See How to Get Started below.

Downstream Use

Integrated into TrialMatch as the eligibility matching step. The app extracts a 13-field patient profile from a PDF, searches a local trial database, and passes each trial through this model to produce verdicts displayed in the Gradio UI.

Out-of-Scope Use

• Not a medical device. This model does not provide medical advice, diagnosis, or treatment decisions.
• Not suitable for final enrollment decisions — it is a pre-screening tool only.
• Not suitable for conditions or trial types not represented in the training data.

---

Bias, Risks, and Limitations

• Training data was synthetically generated using Gemma 4 — the model may reflect biases in that generation process.
• PARTIAL cases are the hardest to distinguish (both misclassifications in the benchmark were PARTIAL ground-truth labels).
• The model has not been validated against real patient records or regulatory enrollment criteria.
• Eligibility criteria text is truncated to 1500 characters — very long criteria sections may lose information.

Recommendations

Use exclusively as a pre-screening aid. Any positive result must be reviewed by a licensed physician before a patient contacts a trial site.

---

How to Get Started with the Model

from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/gemma-4-E4B-it-unsloth-bnb-4bit",
    load_in_4bit=True,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("boweiiismyname/trialmatch-gemma4")
model = PeftModel.from_pretrained(base_model, "boweiiismyname/trialmatch-gemma4")

prompt = """You are a clinical trial eligibility screener.

PATIENT PROFILE:
{patient_profile}

CLINICAL REASONING:
{clinical_reasoning}

ELIGIBILITY CRITERIA:
{eligibility_criteria}

Return exactly:
VERDICT: MATCH|PARTIAL|NO
CONFIDENCE: 0-100
REASON: one sentence
DISQUALIFIERS: criterion text or NONE
NEXT STEP: what to do next"""

inputs = tokenizer(text=prompt.format(
    patient_profile="...",
    clinical_reasoning="...",
    eligibility_criteria="...",
), return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=120, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Training Details

Training Data

120+ synthetic patient profiles generated by Gemma 4 for 60 real recruiting trials sourced from ClinicalTrials.gov. Each profile was run through the base TrialMatch matcher to produce a ground-truth label. Profiles were split roughly 50/50 MATCH-leaning and NO-leaning to prevent class imbalance. Stored as training_data.jsonl.

Training Procedure

Training Hyperparameters

• Training regime: bf16 mixed precision (T4 GPU)
• LoRA rank (r): 16
• LoRA alpha: 16
• Target modules: language layers only (finetune_vision_layers=False)
• Epochs: 3
• Batch size: 2 (effective 8 with gradient accumulation × 4)
• Learning rate: 2e-4, linear schedule with 5 warmup steps
• Optimizer: adamw_8bit
• Max sequence length: 2048

Speeds, Sizes, Times

• Hardware: Google Colab T4 GPU (free tier)
• Training time: ~1.5 hours
• Adapter size: ~50 MB (LoRA weights only)

---

Evaluation

Testing Data

20 held-out examples from training_data.jsonl, sampled with random.seed(99) (different from training split).

Metrics

Exact-match accuracy on VERDICT field (MATCH / PARTIAL / NO). Confidence score is model-reported (0–100).

Results

Both misclassifications were PARTIAL ground-truth — the most ambiguous category. All 3 true MATCH cases were correctly identified, which is the critical metric for a pre-screening tool (missing a match costs a patient trial access).

---

Environmental Impact

• Hardware: NVIDIA T4 (Google Colab)
• Hours used: ~1.5 hours
• Cloud Provider: Google
• Compute Region: us-central1

---

Technical Specifications

Model Architecture

LoRA adapters (r=16) applied to language attention and MLP layers of Gemma 4 E4B. Vision layers are frozen — the adapter is text-only. Base model loaded in 4-bit NF4 quantisation via bitsandbytes.

Software

• unsloth (FastVisionModel)
• trl (SFTTrainer + SFTConfig)
• peft 0.19.1
• transformers
• Python 3.10, CUDA 12.x

---

Model Card Authors

Bowei — built as part of the Gemma 4 Good Hackathon submission.

Model Card Contact

Open an issue on Boweii22/TrialMatch.

Framework versions

• PEFT 0.19.1