ndgold/Qwen3-1.7B-WCAG-AD-4bit

Qwen3-1.7B-WCAG-AD (4-bit, MLX)

LoRA fine-tune of Qwen/Qwen3-1.7B that rewrites the raw Scene: output of the Marlin-2B vision model into one clean, WCAG-compliant audio-description line, for the wcaguar on-device tool. Quantized to 4-bit for Apple-silicon inference via MLX.

It removes meta-reference ("we see", "the camera"), enforces present tense and objectivity (no emotion/intent), drops dialogue echoes, and fits the word budget.

Refiner I/O

• input = Marlin's parsed Scene: text + context (dialogue / previous description / word budget)
• output = one compliant audio-description line within the budget

Evaluation (held-out human-AD test set)

metric	refiner (ft)	raw Marlin
compliance	0.94	0.0813
objectivity	0.9888	0.4425
present tense	0.9938	0.95
no meta-reference	0.9912	0.2487
budget adherence	0.9625	0.675
English LID	1.0	1.0
chrF vs gold	99.59	93.06
ROUGE-L vs gold	0.9951	0.8362

Ship gate: PASS

Usage (MLX)

from mlx_lm import load, generate
model, tok = load("ndgold/Qwen3-1.7B-WCAG-AD-4bit")
msgs = [{"role":"system","content":"<wcaguar AD system prompt>"},
        {"role":"user","content":"Marlin: <raw scene>\nDialogue: (none)\nPrevious: (none)\nBudget: 14 words\nRefined:"}]
p = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False, enable_thinking=False)
print(generate(model, tok, prompt=p, max_tokens=96))

Sources & licenses

• Base model: Qwen3 (Apache-2.0).
• Gold audio descriptions:
  • VideoA11y-40K — WCAG-style audio descriptions (CHI 2025), CC BY-NC-4.0. Primary gold source.
  • You Described, We Archived (YuWA) — human audio-description transcripts from YouDescribe (CC BY-NC-ND 4.0).
  • All gold is filtered against the WCAG / wcaguar audio-description rules: present tense, no meta-reference ("we see"/camera), objectivity (no emotion/intent), no sound description, word budget, English LID.
• Marlin-style drafts: synthesized from the gold AD by reversing the wcaguar style rules (cold-start, denoising SFT). Real Marlin-2B outputs are the documented upgrade path (see DECISIONS.md).
• Intended for the wcaguar on-device WCAG audio-description tool, as a text refiner that rewrites the Marlin-2B vision model's raw Scene: output into a compliant AD line. Non-commercial / academic use (CC BY-NC). Not a general chatbot. No source media is redistributed.