pt-t5-sci-transfer-init-spm32k

PT-Trans-Init

Portuguese scientific T5 model initialized from EN-T5-Sci using WECHSEL and a language-specific SentencePiece 32k tokenizer.

Model Details

This is one of the non-English scientific T5 transfer models from the paper. The model keeps the EN-T5-Sci Transformer weights and reinitializes the language-specific embeddings with WECHSEL using a target SentencePiece tokenizer.

• Paper name: PT-Trans-Init
• Model role: main
• Source/base model: EN-T5-Sci
• Code and pipeline: GitHub repository
• Architecture: T5 encoder-decoder
• SciLaD dataset: scilons/SciLaD-all-text-v1
• Evaluation benchmark: Global-MMLU
• Target-language tokenizer: Portuguese SciLaD split; language-specific SentencePiece 32k tokenizer

Evaluated against:

• PT-Base-CP control: reported as the continued-pretraining control.
• upstream target-language base: reported as the monolingual base comparison.

WECHSEL resources: English fastText embeddings + Portuguese fastText embeddings (pt) with the portuguese bilingual dictionary.

Evaluation

Zero-shot Global-MMLU accuracy reported by the paper aggregation:

Limitations

The model is evaluated primarily with zero-shot Global-MMLU. Downstream task-specific evaluation is recommended before deployment in specialized scientific workflows.

Citation

• Title: Transferring Scientific English Pre-Trained Language Models to Multiple Languages Using Cross-Lingual Transfer
• Authors: Nikolas Rauscher, Fabio Barth, Georg Rehm
• Venue: LREC-COLING 2026, citation details TBA after publication