Barbados road historic handwriting: weights of the 24-model solution
Trained weights of a 24-model handwritten line recognition ensemble built for the Zindi competition Barbados road historic handwriting. The code that trains these models, decodes the test set with them and combines their outputs is the competition code package; this repository holds the weights that package downloads.
No competition data is stored here: no images, no transcriptions and no prediction files.
Layout
| Path | Content |
|---|---|
members/<model>/ |
the trained weights of one ensemble member: a LoRA adapter (adapter/ or model/), a full fine-tune (model/) or CTC weights (model/ or best_weights.pt), with the small records of its training run |
bases/ppocrv6-medium-historical/ |
the PP-OCRv6 medium historical recognizer used as a pretrained base (unmodified files of Zenodo record 21788410, Apache-2.0) |
MANIFEST.json |
every file with its size and SHA-256 |
The package's scripts/download_weights.py mirrors this repository into models/ and verifies every file against
MANIFEST.json.
Models and their pretrained bases
Every model is fine-tuned once on the competition's labeled lines from an open-source pretrained model. Adapters need their base model at load time; the package downloads the bases from their own public repositories at pinned revisions.
Model (members/…) |
Kind | Pretrained base | Licence of the base |
|---|---|---|---|
dots_mocr_ft_vision_hires2_5ep_lr5e6_refit |
full fine-tune | dots-studio/dots.mocr |
MIT |
qwen3_5_lora_vision_refit, qwen3_5_lora_vision_hires2 |
LoRA | Qwen/Qwen3.5-9B |
Apache-2.0 |
qwen3_vl_lora_vision_r64_hires2_s43, qwen3vl_8b_f4 |
LoRA | Qwen/Qwen3-VL-8B-Instruct |
Apache-2.0 |
paddleocr_vl_lora_vision_refit, paddleocr_vl_s11_f1, paddleocr_vl_1_6_e6hi_f4_f4 |
LoRA | PaddlePaddle/PaddleOCR-VL-1.6 |
Apache-2.0 |
lighton_ocr_ft |
full fine-tune | lightonai/LightOnOCR-2-1B-base |
Apache-2.0 |
ppocrv6_htr_ft, ppocrv6_htr_ft_20ep_f0 |
CTC, full fine-tune | PP-OCRv6 medium historical (bases/) |
Apache-2.0 |
dots_mocr_f2, dots_mocr_h112_aug2_f4, dots_mocr_h140_f0 |
LoRA | dots-studio/dots.mocr |
MIT |
firered_ocr_f0 |
LoRA | FireRedTeam/FireRed-OCR |
Apache-2.0 |
glm41v_9b_e2_drop0_f3 |
LoRA | zai-org/GLM-4.1V-9B-Base |
MIT |
gemma26_a4b_f3, gemma26_a4b_f4 |
LoRA | google/gemma-4-26B-A4B-it |
Apache-2.0 |
olmocr2_v2_f1, olmocr2_v2_f3 |
LoRA | allenai/olmOCR-2-7B-1025 |
Apache-2.0 |
convtexth96_f2, armc_f1_s42, convtexts2_f0, convtextd12_f2 |
CTC, full fine-tune | DAIR-Group/HTR-ConvText (checkpoints/iam.pth) |
Apache-2.0 |
All bases are open source and download without an account. The weights in this repository are released under the licence of their base model.
armc_f1_s42, convtexts2_f0 and convtextd12_f2 are produced by the package's run_training.sh; their folders are
added here when that training is finished.
Use
These weights are meant to be used through the competition code package (install.sh downloads them,
run_inference.sh decodes the test set and writes the submission). They are not general-purpose OCR models: they were
trained on a few thousand lines of one archive's seventeenth-century English handwriting.