kauffinger/scibert_scivocab_uncased-mnli

mnli

This model is a fine-tuned version of allenai/scibert_scivocab_uncased on the GLUE MNLI dataset. It achieves the following results on the evaluation set:

• Loss: 0.4917
• Accuracy: 0.8345

Model description

This is the pretrained model presented in SciBERT: A Pretrained Language Model for Scientific Text, which is a BERT model trained on scientific text, then finetuned on GLUE MNLI for zero-shot classification.

The training corpus was papers taken from Semantic Scholar. Corpus size is 1.14M papers, 3.1B tokens. We use the full text of the papers in training, not just abstracts.

SciBERT has its own wordpiece vocabulary (scivocab) that's built to best match the training corpus.

Intended uses & limitations

Zero-shot classification of scientific texts. Note that this model is outperformed by multiple models and was uploaded for research purposes. For actually classifying scientific text, I recommend looking into Deberta v3 Large tuned on MNLI which according to my benchmark on abstracts performs best at current date (7/10/22).

Training and evaluation data

GLUE MNLI

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3.0
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.22.2
• Pytorch 1.11.0+cu113
• Datasets 2.5.1
• Tokenizers 0.12.1

If using these models, please cite the following paper:

@inproceedings{beltagy-etal-2019-scibert,
    title = "SciBERT: A Pretrained Language Model for Scientific Text",
    author = "Beltagy, Iz  and Lo, Kyle  and Cohan, Arman",
    booktitle = "EMNLP",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/D19-1371"
}