OCR-MLDR

An OCR-degraded version of the Multi Long Document Retrieval (MLDR) benchmark, designed to evaluate embedding models on noisy, OCR-like text with long documents.

Dataset Description

A 2,000-document subsample per language was drawn from mteb/MultiLongDocRetrieval (test split). Each passage and query was rendered as a PDF at a specific DPI / font-size setting and re-extracted via OCR using the ocr-robust-multilingual-embeddings OCR simulator to introduce realistic character-level noise.

Both the original clean text and the OCR-noised text are provided, together with the original MLDR relevance judgments (qrels).

Configurations

Config	DPI	Font Size	Description
{lang}_dpi120_font10	120	10 pt	Low quality — high noise
{lang}_dpi130_font10	130	10 pt	Medium quality
{lang}_dpi300_font12	300	12 pt	High quality — low noise

Languages: de, en, es, fr, ru

Configs per Language/DPI

Each (language, DPI) combination provides three configs:

Config	Columns	Description
{lang}_{dpi}_corpus	_id, clean_text, ocr_text	Long document passages
{lang}_{dpi}_queries	_id, clean_text, ocr_text	Search queries
{lang}_{dpi}_qrels	query_id, corpus_id, score	Relevance judgments

All configs use a single split: test.

Sample Sizes (2,000 corpus docs per language)

Language	Corpus	Queries
de	2,000	~200
en	2,000	~800
es	2,000	~200
fr	2,000	~200
ru	2,000	~200

Usage

from datasets import load_dataset

# Load English at DPI 120
corpus  = load_dataset("YOUR_ORG/ocr-mldr", data_dir="data/en_dpi120_font10_corpus",  split="test")
queries = load_dataset("YOUR_ORG/ocr-mldr", data_dir="data/en_dpi120_font10_queries", split="test")
qrels   = load_dataset("YOUR_ORG/ocr-mldr", data_dir="data/en_dpi120_font10_qrels",   split="test")

# Access clean and OCR-noised text
print(corpus[0]["clean_text"])
print(corpus[0]["ocr_text"])

Evaluation

Quick Start (Python)

import numpy as np
from datasets import load_dataset
from sentence_transformers import SentenceTransformer

# 1. Load data
corpus  = load_dataset("YOUR_ORG/ocr-mldr", data_dir="data/en_dpi120_font10_corpus",  split="test")
queries = load_dataset("YOUR_ORG/ocr-mldr", data_dir="data/en_dpi120_font10_queries", split="test")
qrels   = load_dataset("YOUR_ORG/ocr-mldr", data_dir="data/en_dpi120_font10_qrels",   split="test")

# 2. Encode with any SentenceTransformer model
model = SentenceTransformer("Alibaba-NLP/gte-multilingual-base", trust_remote_code=True)
corpus_emb  = model.encode(corpus["ocr_text"],  normalize_embeddings=True, show_progress_bar=True)
query_emb   = model.encode(queries["ocr_text"], normalize_embeddings=True, show_progress_bar=True)

# 3. Retrieve — cosine similarity (embeddings are L2-normalised)
similarities = np.dot(query_emb, corpus_emb.T)

# 4. Compute NDCG@10 for each query
qrels_dict = {}
for row in qrels:
    qrels_dict.setdefault(row["query_id"], {})[row["corpus_id"]] = row["score"]

corpus_ids = corpus["_id"]
for i, qid in enumerate(queries["_id"]):
    if qid not in qrels_dict:
        continue
    top_idx = np.argsort(similarities[i])[::-1][:10]
    hits = [qrels_dict[qid].get(corpus_ids[j], 0) for j in top_idx]
    dcg  = sum(r / np.log2(k + 2) for k, r in enumerate(hits))
    ideal = sorted(hits, reverse=True)
    idcg = sum(r / np.log2(k + 2) for k, r in enumerate(ideal))
    print(f"Query {qid}: NDCG@10 = {dcg / idcg if idcg else 0:.4f}")

Full Evaluation Script

The included evaluation_IR.py evaluates across all languages, DPI settings, and modes in one run:

pip install datasets sentence-transformers torch numpy pandas

python evaluation_IR.py \
    --model Alibaba-NLP/gte-multilingual-base \
    --dataset YOUR_ORG/ocr-mldr \
    --dpi dpi120_font10 dpi130_font10 dpi300_font12 \
    --langs de en es fr ru \
    --mode clean ocr \
    --batch_size 64

Results are saved to ./ir_results/results_latest.csv.

Evaluation Modes

Mode	Corpus	Queries	Purpose
clean	clean_text	clean_text	Upper bound (no OCR noise)
ocr	ocr_text	ocr_text	Realistic full-OCR scenario
clean2ocr	clean_text	ocr_text	Noisy user query against clean index
ocr2clean	ocr_text	clean_text	Clean query against noisy OCR index

Metrics

• NDCG@10 — Normalized Discounted Cumulative Gain
• MRR@10 — Mean Reciprocal Rank
• Recall@100

CER Summary (Character Error Rate)

Mean character-level error rates between clean and OCR-noised text, measured per language and DPI setting.

dpi120_font10 (Low quality)

Language	Corpus CER	Query CER	Corpus Docs	Queries
de	15.4%	8.9%	2,000	200
en	11.0%	9.1%	2,000	800
es	9.4%	6.7%	2,000	200
fr	11.6%	10.1%	2,000	200
ru	12.3%	7.3%	2,000	200

dpi130_font10 (Medium quality)

Language	Corpus CER	Query CER	Corpus Docs	Queries
de	9.9%	5.9%	2,000	200
en	6.7%	4.6%	2,000	800
es	5.1%	3.6%	2,000	200
fr	7.2%	5.1%	2,000	200
ru	8.5%	4.6%	2,000	200

dpi300_font12 (High quality)

Language	Corpus CER	Query CER	Corpus Docs	Queries
de	4.5%	2.6%	2,000	200
en	2.9%	1.1%	2,000	800
es	1.9%	1.2%	2,000	200
fr	3.3%	1.6%	2,000	200
ru	4.9%	1.9%	2,000	200

OCR Noise Generation

Noise was generated with the OCR Simulator using generate_ocr_mldr.py:

1. Texts are split into sentences.
2. Each sentence is rendered as a PDF image (Pillow) at the given DPI / font size.
3. The image is OCR-ed via Tesseract.
4. Sentences are rejoined into documents.

Related

• OCR-MIRACL — Same OCR noise pipeline applied to miracl/miracl.

Citation

If you use this dataset, please cite the OCR noise generation method and MLDR:

@inproceedings{michail-etal-2025-cheap,
    title     = "Cheap Character Noise for {OCR}-Robust Multilingual Embeddings",
    author    = "Michail, Andrianos  and
                 Opitz, Juri  and
                 Wang, Yining  and
                 Meister, Robin  and
                 Sennrich, Rico  and
                 Clematide, Simon",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month     = jul,
    year      = "2025",
    address   = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url       = "https://aclanthology.org/2025.findings-acl.609/",
    pages     = "11705--11716",
    ISBN      = "979-8-89176-256-5"
}

@misc{chen2024bge,
  title={BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
  author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
  year={2024},
  eprint={2402.03216},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}