ClinicalEase-Qwen3-1.7B

LoRA adapter on top of Qwen/Qwen3-1.7B that rewrites clinical / medical text into plain, patient-friendly language. Built to make discharge notes, medication instructions, and provider documentation legible to a non-clinician audience without losing clinical accuracy.

The latest weights live on main. Earlier training phases are preserved as named revisions so the curriculum is reproducible.

Versions

revision	what it is
main / phase2	Latest. Continued training from phase1 on additional / harder data.
phase1	First SFT pass on the base curriculum.
seed20	Earliest run on this task (seed 20). Kept for provenance; superseded by phase1 / phase2.

Load a specific revision with revision="phase1" (or "phase2", "seed20") on PeftModel.from_pretrained.

Usage

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

base_id    = "Qwen/Qwen3-1.7B"
adapter_id = "txmedai/ClinicalEase-Qwen3-1.7B"   # main = phase2
# adapter_id, revision="phase1"  # to pin to an earlier phase

tok  = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

SYSTEM = ("Rewrite the clinical text below in plain, patient-friendly language. "
          "Preserve every clinical fact, dose, and instruction exactly. "
          "Avoid jargon; explain abbreviations; keep it warm and direct.")

def explain(text, max_new_tokens=400):
    msgs = [{"role": "user", "content": f"{SYSTEM}\n\n{text}"}]
    prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
    ids = tok(prompt, return_tensors="pt").to(model.device)
    out = model.generate(**ids, max_new_tokens=max_new_tokens,
                         temperature=0.4, top_p=0.9, do_sample=True,
                         pad_token_id=tok.eos_token_id)
    return tok.decode(out[0][ids.input_ids.shape[1]:], skip_special_tokens=True).strip()

Runs comfortably on a single consumer GPU.

Training

• Framework: TRL SFT
• PEFT: LoRA, r=16, α=32, dropout=0.05, bias=none, task=CAUSAL_LM
• Targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (full attn + MLP)
• Curriculum: seed20 → phase1 → phase2 (continued training on improved data)

Intended use

Generating plain-language explanations of clinical text for patient-facing applications: discharge summaries, medication instructions, lab-result explainers. Not a substitute for a clinician's review. Always verify clinical accuracy before delivering output to a patient.

Limitations

• Small model (1.7 B). Strong rewriting, but limited reasoning over complex multi-system cases.
• Inherits Qwen3-1.7B's training-data biases.
• No safety system for medication-dosage edge cases or contraindications — downstream review is required.

License

Apache-2.0 for the adapter weights. The base model Qwen/Qwen3-1.7B is governed by Qwen3's own license.

History

This repository consolidates three previously-separate repos (ClinicalEase-Qwen3-1.7B, -Phase1, -Phase2). Older repos have been retired in favor of the revision= mechanism above.