mlai-dante/road-model

ROAD Barbados Historic Handwriting Qwen3-VL Adapter

This repository contains a LoRA adapter fine-tuned for the Zindi R.O.A.D. Barbados Historic Handwriting Challenge.

Base model: Qwen/Qwen3-VL-8B-Instruct

The adapter was trained to transcribe cropped images of historical handwritten Barbados archival records into text.

Validation

• cer: 0.079004
• wer: 0.222702
• score: 0.150853

Inference Prompt

You are transcribing a cropped image from an old Barbados archival record.

The image contains handwritten historical text, usually one line or a short phrase from an 18th or 19th century legal, estate, will, deed, or notarial document.

Transcribe only the visible handwritten text.

Important rules:
- Preserve the original wording exactly as written.
- Preserve old, archaic, or unusual spelling. Do not modernize it.
- Preserve capitalization when clear.
- Preserve punctuation when visible.
- Preserve abbreviations, superscript-like marks, contractions, and symbols such as &, y^e, p^rsents, s^d, deced, etc.
- Do not correct grammar.
- Do not expand abbreviations.
- Do not add missing words.
- Do not describe the image.
- Do not explain your answer.
- If a character or word is uncertain, make the best transcription from the visible handwriting.

Return only the transcription text.

The repository includes the fine-tuned LoRA adapter, processor files, and run artifacts such as submission.csv, validation metrics, validation predictions, and the training config. The original challenge images and CSV data are not uploaded.