Topic classification v2.0.0

OpenAlex topic classification, version 2. Since 6 October 2026, every work in OpenAlex gets its topics from this model. The topics themselves are unchanged: same 4,516 topics, same subfields, fields and domains, same IDs. What changed is which works are assigned to each topic.

Start with the v2 README: benchmarks, how the model works, known issues, and how to reproduce every number.

Assets on this release

• topic-classifier-v2__*: the classifier's weights (a Qwen3-8B model with a classification head over the 4,516 topics plus "not classifiable").
• shortlist-student-qwen3-1.7b__* and retriever-e5-base.tar.gz: the shortlisting models used to build the training labels.
• teacher_labels_2m.jsonl.gz, jev_shortlists_first_million.jsonl.gz, retriever_training_labels.jsonl.gz: the training data.
• topics_v1_frozen_part00.csv.gz to topics_v1_frozen_part11.csv.gz: every work's topics under the previous model (version 1), frozen on 5 October 2026, for anyone who needs to reproduce an analysis made with the old assignments. Columns: work_id, topic_1_id, topic_1_score, topic_2_id, topic_2_score, topic_3_id, topic_3_score. Checksums in SHA256SUMS.

Checksums for the models and training data are in v2/models/MANIFEST.json; v2/tools/download.py downloads and verifies them.

Code MIT, data CC0, weights Apache-2.0.