Post

🧵 Excited to share that I'll be presenting SciLaD at #LREC2026! SciLaD: A Large-Scale, Transparent, Reproducible Dataset for Natural Scientific Language Processing arXiv → arxiv.org/abs/2512.1... 1/5

Scientific NLP needs large, open, high-quality data. SciLaD is our answer: a corpus of Open Access publications harvested from Unpaywall, converted into structured TEI XML using open-source tools. 35M+ multilingual publications across languages and scientific domains. 🌍 2/5

We also release a curated English split of 10M+ documents — filtered and deduplicated — along with the full open-source pipeline so anyone can reproduce or extend the dataset. No black boxes. Full transparency. 🔓 3/5

To validate the data quality, we pre-trained SciLaD-M — a RoBERTa encoder trained from scratch on SciLaD. Result: performance comparable to or better than SciBERT across scientific NLP benchmarks, using only openly available data. Open data → open models. 💡 4/5