Abdoul27

dots-mocr-barbados

Local scores

Metric (competition metric): score = 1 − WER_w/24 − CER_w/110, where word and character edit distances are weighted per line by √(reference length). WER/CER columns are plain micro-averaged percentages. Greedy decoding at image scale 1.5 unless noted.

Table with columns: split, lines, score, WER %, CER %, exact lines, role
split	lines	score	WER %	CER %	exact lines	role
dev minus original100 ("dev150")	150	0.9013	14.93	4.56	29	selection set: checkpoint and image scale chosen here
dev250 (dev150 + original100)	250	0.8987	15.38	4.58	46
original100	100	0.8949	16.04	4.61	17	reported once, never used for any choice
audit350	350	0.9004	14.87	4.78	93	holdout, never used for any choice
dev150, zero-shot base model	150	0.7713	32.70	12.45	3	scale 1.5, same usage

Decoding comparison (same checkpoint; beam = 5 beams; "+ char LM" = beam candidates re-ranked with a character 6-gram trained on the training transcriptions, weights λ=0.05, μ=-0.25 chosen on dev150 — the LM is not included in this repo):

Table with columns: split, greedy, beam-5 top-1, beam-5 re-ranked by model log-prob, beam-5 + char LM, best candidate in the 5-best (oracle)
split	greedy	beam-5 top-1	beam-5 re-ranked by model log-prob	beam-5 + char LM	best candidate in the 5-best (oracle)
dev-minus-o100 (tuning set)	0.9013	0.9046	0.9019	0.9042	0.9260
audit350 (holdout)	0.9004	0.9033	0.9026	0.9028	0.9291
original100 (once)

Paired bootstrap on audit350: beam-5 vs greedy +0.0022 (95% CI -0.0012 to +0.0067); beam-5 + char LM vs greedy +0.0024 (95% CI -0.0015 to +0.0069).

Data splits (frozen, SHA-256 verified images)

Table with columns: split, lines, use
split	lines	use
train3497	3,497	training only
dev150 = dev250 minus original100	150	every choice: checkpoint, image scale, decoding settings
original100	100	reported once at the end; never used for any choice
audit350	350	holdout for comparisons; never tuned on
test	1,374	competition test lines (no labels)

The splits are disjoint by line ID and by exact image hash.

Requirements

transformers==4.57.6 (the base repo's pin; the model uses remote code: trust_remote_code=True)
torch>=2.4 with torchvision, Pillow, huggingface_hub, accelerate
a CUDA GPU with bf16 (≈ 14 GB for inference: the weights are FP32, ~12 GB). flash-attn is optional: without it barbados_inference.py registers a stub and uses PyTorch SDPA attention (same results).

bash

pip install "transformers==4.57.6" torch torchvision pillow huggingface_hub acceleratehuggingface-cli login        # after your access request is approved

How to use

python

from huggingface_hub import snapshot_downloadimport syspath = snapshot_download("Abdoul27/dots-mocr-barbados")sys.path.insert(0, path)from barbados_inference import load, transcribemodel, processor = load(path)print(transcribe(model, processor, ["line_001.jpg", "line_002.jpg"]))            # greedy (scores above)print(transcribe(model, processor, ["line_001.jpg"], num_beams=5))               # beam-5 (see decoding comparison)

What barbados_inference.py does (reproduce this exactly to get the scores above):

Input: one image per handwritten line (not a full page), converted to RGB and resized by 1.5× (bicubic), capped at 1,605,632 pixels.
Prompt: the official dots.mocr OCR prompt "Extract the text content from this image.", image first, no system prompt, through the model's own chat template.
Decoding: greedy (or beam search), at most 51 new tokens, stop at <|endofassistant|> (also <|endoftext|>, <|assistant|>); FP32 weights with BF16 autocast; one image per call. Vision attention runs per image with PyTorch SDPA, and beam search copies the image correctly for every beam (the default expansion would scramble dots.mocr's flat patch matrix).
Output: multi-line outputs are joined with single spaces; repetition loops are trimmed only when the length cap is hit.

Transcription conventions learned from the training labels: original spelling and abbreviations are kept (pnts, Xpian, w^th); ^ marks raised (superscript) letters, e.g. w^th, Exec:^rs; & is kept.

Training

Full fine-tune of all weights of dots-studio/dots.mocr: FP32 weights with BF16 autocast, AdamW (β = (0.9, 0.95), weight decay 0.1 except norms/biases/embeddings; 8-bit AdamW on GPUs under 70 GB), lr 1e-05 (vision encoder 2e-06), warm-up then cosine to 10 %, effective batch 16, up to 5 epochs with evaluation twice per epoch; best checkpoint epoch_05_half, selected on dev150.
Answer-only loss including the end token <|endofassistant|>; mild augmentation (±10 % scale, ±1.5° rotation, brightness/contrast, light blur); gradient checkpointing. barbados_training_worker.py is the complete training/evaluation code.

Limitations

Line-level model: it expects one cropped text line per image; full pages need line segmentation first.
Trained on 3497 lines from one archive and period; other hands, languages or scripts may degrade.

Licence

Same licence as the base model dots-studio/dots.mocr (mit; see LICENSE when present).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Abdoul27

Model Tree

Base

dots-studio/dots.mocr

Fine-tuned

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Local scores

Table with columns: split, lines, score, WER %, CER %, exact lines, role
split	lines	score	WER %	CER %	exact lines	role
dev minus original100 ("dev150")	150	0.9013	14.93	4.56	29	selection set: checkpoint and image scale chosen here
dev250 (dev150 + original100)	250	0.8987	15.38	4.58	46
original100	100	0.8949	16.04	4.61	17	reported once, never used for any choice
audit350	350	0.9004	14.87	4.78	93	holdout, never used for any choice
dev150, zero-shot base model	150	0.7713	32.70	12.45	3	scale 1.5, same usage

Table with columns: split, greedy, beam-5 top-1, beam-5 re-ranked by model log-prob, beam-5 + char LM, best candidate in the 5-best (oracle)
split	greedy	beam-5 top-1	beam-5 re-ranked by model log-prob	beam-5 + char LM	best candidate in the 5-best (oracle)
dev-minus-o100 (tuning set)	0.9013	0.9046	0.9019	0.9042	0.9260
audit350 (holdout)	0.9004	0.9033	0.9026	0.9028	0.9291
original100 (once)

Paired bootstrap on audit350: beam-5 vs greedy +0.0022 (95% CI -0.0012 to +0.0067); beam-5 + char LM vs greedy +0.0024 (95% CI -0.0015 to +0.0069).

Data splits (frozen, SHA-256 verified images)

Table with columns: split, lines, use
split	lines	use
train3497	3,497	training only
dev150 = dev250 minus original100	150	every choice: checkpoint, image scale, decoding settings
original100	100	reported once at the end; never used for any choice
audit350	350	holdout for comparisons; never tuned on
test	1,374	competition test lines (no labels)

The splits are disjoint by line ID and by exact image hash.

Requirements

transformers==4.57.6 (the base repo's pin; the model uses remote code: trust_remote_code=True)
torch>=2.4 with torchvision, Pillow, huggingface_hub, accelerate
a CUDA GPU with bf16 (≈ 14 GB for inference: the weights are FP32, ~12 GB). flash-attn is optional: without it barbados_inference.py registers a stub and uses PyTorch SDPA attention (same results).

bash

pip install "transformers==4.57.6" torch torchvision pillow huggingface_hub acceleratehuggingface-cli login        # after your access request is approved

How to use

python

from huggingface_hub import snapshot_downloadimport syspath = snapshot_download("Abdoul27/dots-mocr-barbados")sys.path.insert(0, path)from barbados_inference import load, transcribemodel, processor = load(path)print(transcribe(model, processor, ["line_001.jpg", "line_002.jpg"]))            # greedy (scores above)print(transcribe(model, processor, ["line_001.jpg"], num_beams=5))               # beam-5 (see decoding comparison)

What barbados_inference.py does (reproduce this exactly to get the scores above):

Input: one image per handwritten line (not a full page), converted to RGB and resized by 1.5× (bicubic), capped at 1,605,632 pixels.
Prompt: the official dots.mocr OCR prompt "Extract the text content from this image.", image first, no system prompt, through the model's own chat template.
Decoding: greedy (or beam search), at most 51 new tokens, stop at <|endofassistant|> (also <|endoftext|>, <|assistant|>); FP32 weights with BF16 autocast; one image per call. Vision attention runs per image with PyTorch SDPA, and beam search copies the image correctly for every beam (the default expansion would scramble dots.mocr's flat patch matrix).
Output: multi-line outputs are joined with single spaces; repetition loops are trimmed only when the length cap is hit.

Training

Full fine-tune of all weights of dots-studio/dots.mocr: FP32 weights with BF16 autocast, AdamW (β = (0.9, 0.95), weight decay 0.1 except norms/biases/embeddings; 8-bit AdamW on GPUs under 70 GB), lr 1e-05 (vision encoder 2e-06), warm-up then cosine to 10 %, effective batch 16, up to 5 epochs with evaluation twice per epoch; best checkpoint epoch_05_half, selected on dev150.
Answer-only loss including the end token <|endofassistant|>; mild augmentation (±10 % scale, ±1.5° rotation, brightness/contrast, light blur); gradient checkpointing. barbados_training_worker.py is the complete training/evaluation code.

Limitations

Line-level model: it expects one cropped text line per image; full pages need line segmentation first.
Trained on 3497 lines from one archive and period; other hands, languages or scripts may degrade.

Licence

Same licence as the base model dots-studio/dots.mocr (mit; see LICENSE when present).

dots-mocr-barbados

README

Local scores

Data splits (frozen, SHA-256 verified images)

Requirements

How to use

Training

Limitations

Licence

Explore FriendliAI today

README

Local scores

Data splits (frozen, SHA-256 verified images)

Requirements

How to use

Training

Limitations

Licence