Datasets:

jeffdekerj
/

bms-images-shards

license: cc0-1.0
task_categories:
  - image-to-text
tags:
  - chemistry
  - molecular-structure
  - smiles
  - ocr
  - computer-vision
  - webdataset
  - lightonocr
size_categories:
  - 1M<n<10M

BMS Molecular Translation - WebDataset Shards

This dataset contains pre-processed WebDataset shards of the BMS Molecular Translation dataset, optimized for fast data loading during model training.

Dataset Summary

Total Size: 3.8 GB
Training shards: 236 files (3.7 GB) - 2.36M molecular structure images with SMILES
Validation shards: 5 files (0.1 GB) - 48K samples for model validation
Test shards: 3 files (0.0 GB) - 24K held-out samples for final evaluation

Format

Shards are in WebDataset format:

Sequential tar archives for fast I/O
10,000 samples per shard
Training data pre-shuffled
Val/test data in original order
Tar files are preserved (not extracted) - perfect for WebDataset!

Usage

Download the Dataset

# Using HuggingFace Hub
pip install huggingface_hub

# Download entire dataset
python download_shards_from_huggingface.py --username jeffdekerj

# Or use HuggingFace Hub directly
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="jeffdekerj/bms-images-shards",
    repo_type="dataset",
    local_dir=".data/webdataset_shards"
)

Load with WebDataset

from webdataset_loader import BMSWebDataset
from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("lightonai/LightOnOCR-1B-1025")

train_dataset = BMSWebDataset(
    shard_dir=".data/webdataset_shards/train/",
    processor=processor,
    user_prompt="Return the SMILES string for this molecule.",
    shuffle_buffer=1000,
)

Train Your Model

python finetune_lightocr.py \
  --train_shards .data/webdataset_shards/train/ \
  --val_shards .data/webdataset_shards/val/ \
  --per_device_train_batch_size 4 \
  --num_train_epochs 3 \
  --fp16

Benefits

2-5x faster data loading vs individual files
Better I/O performance for network filesystems
Lower overhead with sequential reads
Built-in shuffling without memory overhead
Tar files preserved - no auto-extraction like Kaggle

Source Repository

GitHub: https://github.com/JeffDeKerj/lightocr

Complete documentation available in the repository:

docs/WEBDATASET_GUIDE.md - Complete usage guide
docs/HUGGINGFACE_GUIDE.md - HuggingFace-specific guide
docs/FINETUNE_GUIDE.md - Fine-tuning guide
README.md - Project overview

Original Dataset

Based on the BMS Molecular Translation competition dataset: https://www.kaggle.com/c/bms-molecular-translation