Lumen estimated ~200× lower inference cost with frontier competitive saferepair

Lumen is a 4b model for automatically verifiable accessibility attribute repair in react and static html. on a matched 43-task development benchmark, Lumen completed 42 safe repairs, compared with 40 for gpt-5.6 sol and 43 for claude opus 4.8.

this repository is the lightweight 252 mb lora adapter edition of Lumen. it requires the exact qwen3.5 4b base checkpoint shown below.

most users should use the standalone harrrshall/lumen-merged repository instead. it is approximately 9.1 gb, loads directly with transformers, and does not require attaching a separate adapter.

quick start for this adapter

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base_id = "Qwen/Qwen3.5-4B"
base_revision = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"

processor = AutoProcessor.from_pretrained(base_id, revision=base_revision)
base = AutoModelForImageTextToText.from_pretrained(
    base_id,
    revision=base_revision,
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "harrrshall/lumen")
model.eval()

do not download both editions. choose lumen-merged for a simple standalone setup, or this adapter when you already store the pinned base model and want the smaller download.

Lumen generates small structured patches for automatically verifiable accessibility defects in react and static html source code.

model details

field	value
model name	Lumen
base model	qwen3.5 4b
training method	qlora supervised fine tuning
working context	16,384 tokens
output	structured source patch
license	apache 2.0

frontier comparison

system	tasks	saferepair	cost per safe repair
Lumen	43 tasks	42/43 (97.67%)	$0.000268
gpt-5.6 sol	43 tasks	40/43 (93.02%)	$0.070857
claude opus 4.8	43 tasks	43/43 (100%)	$0.054012

evaluation gate

browser saferepair requires target findings to be removed without new findings, plus passing builds, selected tests, keyboard and focus checks, semantic and functional invariants, visual policy, protected path checks, and evaluator evasion controls.

training data

the current synthetic corpus contains 637 verified training tasks, 100 development tasks, and 100 sealed synthetic tasks. the trajectory set contains 30,576 training records and 100 development records. a11ybench is reserved for evaluation and excluded from training.

training curve

intended use

• accessibility repair assistance in code review
• validated patch generation for continuous integration
• scoped remediation for react and static html repositories
• private deployment for source sensitive workflows

limitations

Lumen covers deterministic source repair families supported by its evaluator. it does not establish complete wcag or legal conformance. review every patch before merging, and involve qualified accessibility practitioners and assistive technology users for subjective decisions.