en-t5-sci-continued-pretraining-487k

EN-T5-Sci

Continued-pretrained T5-base on a cleaned English scientific corpus derived from Unpaywall.
Checkpoint: pretraining_logs_lr_001_OPTIMIZED_clean_restart/.../step-487500-val_ppl-3.72168.ckpt (see conversion_info.json for provenance).

Model Details

• Architecture: T5-base (12L / 12L, d_model=768, 220M params)
• Objective: Span corruption (15 % noise, mean span length 3)
• Sequence prep: Sliding windows of 512 tokens with 50 % overlap
• Optimizer: Adafactor with linear warmup (20k) → inverse sqrt decay, lr=1e-3, grad clip=1.0
• Hardware: 4× NVIDIA H100 (mixed precision, gradient accumulation 2, effective batch 384)

Training Data

English scientific text (approx. 230 GB, ~11 M docs) cleaned with DataTrove and custom regex rules (see thesis Section “Automatic Data Preprocessing”). Tokenization via SentencePiece (original T5 vocab).

Evaluation (Global-MMLU, zero-shot, Global benchmark)

Full plots + per-subtask CSV: evaluation_results/scientific_crosslingual_transfer_eval_full_15k/.

Intended Use

Zero-shot scientific QA, warm-start for downstream fine-tuning on English scientific NLP tasks. Use T5ForConditionalGeneration.from_pretrained("rausch/en-t5-sci-continued-pretraining-487k").

Limitations

• Same T5-base context length (512) and tokenization.
• Evaluated only on Global-MMLU EN/DE; other tasks may require finetuning.
• Training corpus is English-only; no guarantees about other languages.

Citation

Please cite the Bachelor’s thesis (link) and Raffel et al. (2020) for T5.