Dataset Viewer
The dataset viewer is not available for this dataset.
Cannot get the config names for the dataset.
Error code:   ConfigNamesError
Exception:    ValueError
Message:      
Expected data_files in YAML to be either a string or a list of strings
or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'mlp-train', 'path': 'mlp-train-*.parquet'}, {'split': 'full-train', 'path': 'full-train-*.parquet'}, {'split': 'sft-train', 'path': 'sft-train-*.parquet'}, {'split': 'dev', 'path': 'dev-*.parquet'}, {'split': 'test', 'path': 'test-*.parquet'}]
Examples of data_files in YAML:

   data_files: data.csv

   data_files: data/*.png

   data_files:
    - part0/*
    - part1/*

   data_files:
    - split: train
      path: train/*
    - split: test
      path: test/*

   data_files:
    - split: train
      path:
      - train/part1/*
      - train/part2/*
    - split: test
      path: test/*

PS: some symbols like dashes '-' are not allowed in split names

Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 66, in compute_config_names_response
                  config_names = get_dataset_config_names(
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                                   ^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1207, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1182, in dataset_module_factory
                  ).get_module()
                    ^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 611, in get_module
                  metadata_configs = MetadataConfigs.from_dataset_card_data(dataset_card_data)
                                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/utils/metadata.py", line 153, in from_dataset_card_data
                  cls._raise_if_data_files_field_not_valid(metadata_config)
                File "/usr/local/lib/python3.12/site-packages/datasets/utils/metadata.py", line 100, in _raise_if_data_files_field_not_valid
                  raise ValueError(yaml_error_message)
              ValueError: 
              Expected data_files in YAML to be either a string or a list of strings
              or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'mlp-train', 'path': 'mlp-train-*.parquet'}, {'split': 'full-train', 'path': 'full-train-*.parquet'}, {'split': 'sft-train', 'path': 'sft-train-*.parquet'}, {'split': 'dev', 'path': 'dev-*.parquet'}, {'split': 'test', 'path': 'test-*.parquet'}]
              Examples of data_files in YAML:
              
                 data_files: data.csv
              
                 data_files: data/*.png
              
                 data_files:
                  - part0/*
                  - part1/*
              
                 data_files:
                  - split: train
                    path: train/*
                  - split: test
                    path: test/*
              
                 data_files:
                  - split: train
                    path:
                    - train/part1/*
                    - train/part2/*
                  - split: test
                    path: test/*
              
              PS: some symbols like dashes '-' are not allowed in split names

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

LaTeX OCR Dataset

Dataset for training LaTeX OCR models — converting images of mathematical formulas to LaTeX source. Built with a 3-stage curriculum training pipeline.

Sources

Splits

Split Samples Shards Size Description
mlp-train 574,490 7 3.1 GB Stage 1 — light augmentation (blur, rotation, resolution resample)
full-train 127,108 3 1.1 GB Stage 2 — heavy augmentation (JPEG, noise, perspective, erase)
sft-train 2,416 1 20 MB Stage 3 — handwriting data for SFT
dev 16,950 1 76 MB Validation — raw images, no augmentation
test 16,557 1 75 MB Test — raw images, no augmentation

Format

Each row contains:

Column Type Description
index int64 Global sample index
image Image Formula image (JPEG, arbitrary resolution)
label string LaTeX formula string

Label Statistics

Split Mean tokens Median tokens P95 tokens Vocab size
mlp-train 65.0 55 147 1,281
full-train 60.5 53 128 821
sft-train 55.5 46 126 372
dev 62.8 54 137 602
test 62.6 54 136 601

Token count uses the regex \\[a-zA-Z]+|[^\s]. All samples are filtered to [2, 200] tokens.

Augmentation Pipeline

Stage 1 — Light (mlp-train)

  • Inception crop, multi-scale resolution resample (192–768px long side)
  • Gaussian blur (p=0.3), rotation ±3° (p=0.3)
  • Token drop (patch masking, decaying probability)

Stage 2 — Heavy (full-train)

  • Inception crop, multi-scale resolution resample
  • JPEG compression quality 30–75 (p=0.4), Gaussian noise (p=0.3), perspective distortion (p=0.3)
  • Token drop

Stage 3 — SFT (sft-train)

  • Handwriting data (synthetic + human) with heavy augmentation
  • Replay subset from stage 2 to prevent catastrophic forgetting

Dev / Test

  • Raw images, no augmentation

Usage

from datasets import load_dataset

ds = load_dataset("<repo_id>")

# Stage 1 training
for sample in ds["mlp-train"]:
    image = sample["image"]   # PIL.Image
    label = sample["label"]   # str, e.g. "\\frac{1}{2}"
    index = sample["index"]   # int

# Evaluation
for sample in ds["dev"]:
    ...

Training Notes

  • Images have arbitrary resolution — designed for NaViT-style models that handle variable-size inputs natively
  • resize_image() should be called in the dataloader if using a fixed-resolution encoder
  • Tokenizer: Qwen/Qwen2.5-Coder-1.5B — verified full coverage, no UNK tokens
  • Recommended training order: mlp-trainfull-trainsft-train
Downloads last month
51