Clinical Text Simplifier · LoRA Fine-Tuning Case Study

Simpler Isn't Always Safer

Can an open, 0.5B-parameter LLM be LoRA-tuned on a CPU to simplify biomedical text better than the same model prompted zero-shot? Scored against PLABA — NIH/NLM's real, professionally human-simplified dataset — not asserted.

+5.06SARI improvement over baseline
+83%relative sacreBLEU improvement
0.44%of model parameters trained (LoRA)
01

Clinical writing is dense on purpose — can a tiny model fix that?

Large frontier models can simplify biomedical text reasonably well out of the box. That isn't the interesting question for a portfolio project — the narrower, falsifiable one is: can a small (~0.5B parameter) open model, fine-tuned with LoRA on a CPU in a modest amount of wall-clock time, get measurably better at this one task than the same base model prompted zero-shot?

Scored against PLABA — 75 consumer-health topics, 749 PubMed abstracts, every sentence adapted into plain language by professional NLM annotators. Real human references, not self-graded output.

02

CPU-only, validated before committing

This ran on a 4-core laptop CPU, no GPU — the same class of hardware that caused an out-of-memory crash on a past project. A smoke test came back clean, but at 8.8s/example, three epochs over the full 6,202-example split would have taken ~45 hours.

Model

Qwen2.5-0.5B-Instruct

Apache 2.0, 496M params including the LM head.

Method

LoRA, rank 16

2.16M trainable params — 0.44% of the full model. Base weights never change.

Throughput

batch_size=1 fastest

Larger batches and extra threads were both slower on this CPU — benchmarked, not assumed.

Real run

800 examples, 1 epoch

~70–90 minutes — a documented scope reduction, not a hidden one.

03

Every metric moved the same direction

200 held-out test sentences, identical greedy decoding for both models, scored against the same human references.

SARI+5.06
Zero-shot
29.04
Fine-tuned
34.11
sacreBLEU+83%
Zero-shot
9.48
Fine-tuned
17.40
ROUGE-L+0.084
Zero-shot
0.295
Fine-tuned
0.380

Readability moved further too: fine-tuned outputs average Flesch-Kincaid grade 11.08 vs. 12.24 zero-shot, against source sentences averaging grade 14.6.

04

Reading the wins

Clean win
Source
Currently, only one half of the people who have diabetes mellitus have been diagnosed.
Reference
Currently, only half of those with diabetes mellitus have been identified.
Baseline
Currently, less than half of individuals suffering from type 2 diabetes have been identified as having the condition.
Fine-tuned
Only half of those with diabetes have been diagnosed.
Modest, real improvement
Source
These changes provide an easier and more reliable means of diagnosing persons at risk of complications from hyperglycemia.
Reference
These changes help improve identification of those at risk from effects of high blood sugar.
Baseline
These adjustments make it easier to identify people who might have problems with high blood sugar earlier on.
Fine-tuned
These changes make it easier to diagnose people who may have problems with high blood sugar.
05

An honest failure

The fine-tune helped on every metric — that isn't the same claim as "safe for real patient-facing text." Reported directly rather than smoothed over:

Content dropped
Source
It is clear that these changes in diagnostic criteria have major importance both for individuals and for resource planning at a national level.
Reference
It is clear that these changes in identification criteria have major importance both for individuals and for planning at a national level.
Baseline
These changes in diagnostic criteria are crucial for both individual patients and overall resource management at a national scale.
Fine-tuned
It's important to change how doctors diagnose patients so they can better treat them.
A separate output truncates and appears to invent a numeric value ("…lowered from 140 mg/dL to 12") that isn't in the source sentence at all — exactly the kind of faithfulness failure automatic SARI/BLEU/ROUGE metrics are not designed to catch.
06

What surprised us

Given how little supervision 800 examples/1 epoch represents, a split result across metrics seemed more likely going in — better BLEU/ROUGE from picking up PLABA's vocabulary, but no real readability gain, or vice versa. Instead every metric moved the same direction, by a consistent margin.

The fine-tuned model also simplifies more aggressively than the human annotators do by grade level (−3.51 grades vs. source, versus the human references' own −1.40) — and that isn't a free win. Some of that extra simplification is genuinely tighter phrasing; some of it is the model dropping real content for vaguer language that still scores well on n-gram overlap.

07

Limitations

The official TREC-2024 test split sits behind NIST registration that couldn't be completed non-interactively — this project uses a self-constructed topic-level split instead, so numbers aren't directly comparable to the published leaderboard. A single 0.5B model, one LoRA configuration, no hyperparameter sweep. Automatic metrics only — no clinician or lay-reader panel reviewed outputs. Evaluation used 200 of 1,242 test sentences, a documented scope reduction for CPU wall-clock time.