---
dataset_info:
  features:
  - name: synthetic_wrong_text
    dtype: string
  - name: text
    dtype: string
  - name: cer
    dtype: float64
  splits:
  - name: train
    num_bytes: 10857034
    num_examples: 94909
  download_size: 7622382
  dataset_size: 10857034
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

Interpres OCR Correction (IOC) is the dataset that is desgined for training ByT5 models to correct transcribed Medieval Latin texts. The data is collected from the OCR results from multiple OCR models/engines like [Deepseek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR), [PaddleOCR-v5](https://huggingface.co/collections/PaddlePaddle/pp-ocrv5), [TrOCR-Medieval](https://huggingface.co/collections/medieval-data/trocr-medieval-htr), [TRIDIS](https://huggingface.co/magistermilitum/tridis_v2_HTR_historical_manuscripts), ... on the [CATMuS](https://huggingface.co/datasets/CATMuS/medieval) Latin dataset. The CER range is 0.1-0.4 to make sure the dataset doesn't capture too much gibberish cases and also not too easy.