LightOnOCR-2-1B-catmus-textualis

LightOnOCR-2-1B for Latin (Line-Level)

<p align="center"> <img src="https://huggingface.co/lightonai/LightOnOCR-2-1B-base/resolve/main/lightonocr-banner.png" alt="LightOnOCR Banner" width="600"/> </p>

This model is a fine-tuned version of lightonai/LightOnOCR-2-1B-base specifically trained for line-level OCR.

CATMuS Textualis OCR model trained on line-level images from late medieval Gothic manuscripts.

Model Description

• Base Model: lightonai/LightOnOCR-2-1B-base
• Training Data: wjbmattingly/catmus-edited2
• Task: Line-level text transcription from document images
• Language: Latin (la)
• Architecture: Vision-Language Model (1B parameters)

This is a line-level model - it expects cropped line images as input, not full pages. Each image should contain a single line of text.

Evaluation Results

Evaluated on 50 samples from the test set:

Lower CER/WER is better. Higher perfect matches is better.

Example Outputs

✓ = exact match

Usage

Installation

# Requires transformers from source
pip install git+https://github.com/huggingface/transformers
pip install pillow torch

Python Usage

import torch
from transformers import LightOnOcrForConditionalGeneration, LightOnOcrProcessor
from PIL import Image

# Load model and processor
model_id = "wjbmattingly/LightOnOCR-2-1B-catmus-textualis"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if device == "cuda" else torch.float32

processor = LightOnOcrProcessor.from_pretrained(model_id)
model = LightOnOcrForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=dtype,
).to(device)

# Load your line image
image = Image.open("your_image.jpg").convert("RGB")

# Prepare input
messages = [{"role": "user", "content": [{"type": "image"}]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = processor(
    text=[text],
    images=[[image]],
    return_tensors="pt",
    padding=True,
    size={"longest_edge": 700},
).to(device)
inputs["pixel_values"] = inputs["pixel_values"].to(dtype)

# Generate transcription
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)

# Decode output
input_length = inputs["input_ids"].shape[1]
generated_ids = outputs[0, input_length:]
transcription = processor.decode(generated_ids, skip_special_tokens=True)

print(transcription)

Batch Inference

from datasets import load_dataset

# Load dataset
dataset = load_dataset("wjbmattingly/catmus-edited2", split="train[:10]")

# Process batch
images = [[img.convert("RGB")] for img in dataset["image"]]
messages = [{"role": "user", "content": [{"type": "image"}]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
texts = [text] * len(images)

inputs = processor(
    text=texts,
    images=images,
    return_tensors="pt",
    padding=True,
    size={"longest_edge": 700},
).to(device)
inputs["pixel_values"] = inputs["pixel_values"].to(dtype)

outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
predictions = processor.batch_decode(outputs[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)

for pred, gt in zip(predictions, dataset["text"]):
    print(f"Prediction: {pred}")
    print(f"Ground Truth: {gt}")
    print()

Training Details

• Base Model: lightonai/LightOnOCR-2-1B-base
• Training Method: Fine-tuning with frozen language model backbone
• Optimizer: AdamW (fused)
• Learning Rate: 6e-5 with linear decay
• Precision: bfloat16

Limitations

• This model is trained on line-level images. For full-page transcription, you need to first segment the page into individual lines.
• Performance may vary on document styles not represented in the training data.

Citation

If you use this model, please cite:

@misc{lightonocr2_finetuned_2026,
  title = {LightOnOCR Fine-tuned for Latin},
  author = {William Mattingly},
  year = {2026},
  howpublished = {\url{https://huggingface.co/wjbmattingly/LightOnOCR-2-1B-catmus-textualis}}
}

And the original LightOnOCR paper:

@misc{lightonocr2_2026,
  title = {LightOnOCR: A 1B End-to-End Multilingual Vision-Language Model for State-of-the-Art OCR},
  author = {Said Taghadouini and Adrien Cavaill\`{e}s and Baptiste Aubertin},
  year = {2026},
  howpublished = {\url{https://arxiv.org/pdf/2601.14251}}
}

Acknowledgments

• LightOn AI for the excellent LightOnOCR base model
• The creators of the wjbmattingly/catmus-edited2 dataset