- Add `environment.yml` to simplify installation with Conda - Bugfix: force QIDs in folds.json files to be strings - Bugfix: add `multithreaded` option to Pytorch trainer controlling whether a separate worker is used with Pytorch `DataLoader`. Default to false; true is faster, but conflicting Python packages appear to cause deadlock. This is reproducible on Colab and when using Anaconda's default packages (rather than installing from Miniconda). - Update docs
0.2.4.1
Bug fix: allow spaces in paths passed to Anserini (e.g., the collection path)
0.2.4
- New rerankers: PARADE and BERT-MaxP variants - Collection iterator: `for doc in collection_object: ...` - Experimental support for Birch reranker - Bug fix: `evaluator` import should no longer break Python 3.6
0.2.3
- Report results interpolated with searcher - Add new, more efficient `EmbedText` that avoids generating random embeddings for OOV terms. The previous extractor is now `SlowEmbedText` - Bug fix: Add missing NF queries file
0.2.2
- New [NFCorpus](https://www.cl.uni-heidelberg.de/statnlpgroup/nfcorpus/) Collection and Benchmark - Module cleanup - Documentation improvements
0.2.1
- Searcher refactor, with BM25, BM25RM3, BM25PRF, AxiomaticSemanticMatching, DirichletQL, and QLJM supporting lists of parameters - support setting a relevance level threshold (i.e., `trec_eval --level_for_rel`) - miscellaneous cleanup and convenience methods