TrialMatch Qwen3-1.7B LoRA (Run A, selected on validation)

LoRA adapter (r=16, alpha=32, all attention + MLP projections) for Qwen/Qwen3-1.7B, fine-tuned for criterion-level clinical trial eligibility: input = numbered patient note + criterion type + criterion text, output = {"evidence": [sentence IDs], "label": "<label>"}.

Educational screening prototype for a university assignment (GenAI Assignment 2). Not a clinical enrollment decision tool.

  • Data: ncbi/TrialGPT-Criterion-Annotations, patient-grouped split (37/8/8 patients, numpy default_rng(42)); GPT-4 columns never used.
  • Training: answer-token-only loss, 3 epochs, lr 2e-4 cosine, effective batch 16, fp16 on a Kaggle T4; best checkpoint by validation loss.
  • Held-out test (206 criteria, greedy): accuracy 0.748 (base model 0.335), macro-F1 0.571 (base 0.270), evidence micro-F1 0.494 (base 0.416).
  • Code, split manifest and full results: https://github.com/AfnanRizwan1/GenAi-A2
from transformers import AutoModelForCausalLM
from peft import PeftModel
model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B"), "afnanrizwan/trialmatch-qwen3-1.7b-lora").merge_and_unload()
Downloads last month
27
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for afnanrizwan/trialmatch-qwen3-1.7b-lora

Finetuned
Qwen/Qwen3-1.7B
Adapter
(950)
this model

Dataset used to train afnanrizwan/trialmatch-qwen3-1.7b-lora