Ilia-Iliev/arxiv-abstract-classifier-bert-base-uncased

arXiv abstract classifier — BERT base uncased

A PyTorch multilabel classifier for five broad arXiv areas: biology, chemistry, computer science, physics, and social sciences. It was full-fine-tuned from `google-bert/bert-base-uncased` on normalized arXiv abstracts. Inputs are abstracts only; labels are ordered as biology, chemistry, computer_science, physics, and social_sciences.

Evaluation

Results are from the untouched 20,000-record final holdout split. Per-label thresholds were selected on a separate 20,000-record validation split.

The repository contains the tokenizer, backbone (model.safetensors), PyTorch classification head (classifier.pt), thresholds and data/model configuration. Load it with this project's classifier.modeling implementation; the head is not an AutoModel classification head. Use the preprocessing version recorded in configuration.json, including its <FORMULA> placeholder handling.

The benchmark protocol and full predictions/metrics are archived with the project release. Bootstrap uncertainty is over held-out examples, not random training seeds.