Barbados HTR — fine-tuned weights (wahaym, R.O.A.D. Barbados Historic Handwriting Challenge)
Weights for the code-review package of the solution by wahaym to the Zindi R.O.A.D. Barbados
Historic Handwriting Challenge (line-level handwriting recognition of 17th–18th-century Barbadian
deed / protest registers). They are downloaded by scripts/download_weights.py of that package
(called by install.sh) into its models/ folder; they are not meant to be used on their own.
| folder | content |
|---|---|
checkpoints/<model>/ (66) |
fine-tuned models of the ensemble: LoRA adapters (adapter_model.safetensors + adapter_config.json + tokenizer/processor files) or, for full fine-tunes, last.pt (params + buffers) |
checkpoints/PROVENANCE.json |
the training run and epoch each folder was exported from |
weights/party, weights/ppocrv6_medium, weights/satrn |
unmodified public pretrained bases: party (Zenodo 20642057, Apache-2.0), PP-OCRv6 medium (Zenodo 21788410, Apache-2.0), SATRN (Riksarkivet/satrn_htr, MIT) |
The fine-tuned weights were trained on the competition data only and sit on these public base models, whose licences apply: zai-org/GLM-OCR (MIT); Qwen/Qwen3-VL-4B/8B/32B-Instruct, PaddlePaddle/PaddleOCR-VL-1.6, ATH-MaaS/OvisOCR2, lightonai/LightOnOCR-2-1B-base, google/gemma-4-E4B-it, Kansallisarkisto/multicentury-htr-model, wjbmattingly/nara-qwen-3.5-2b (Apache-2.0); ENC-PSL/Medusa0.1Line-4B (CC-BY-4.0).