S2AND

S2AND provides author-name-disambiguation datasets and the reference model from S2AND: A Benchmark and Evaluation System for Author Name Disambiguation. The model predicts author clusters from paper and author records.

This branch prepares Python/Rust runtime 1.0.0 and model/data release 1.3.

Update at v1.3 release: Runtime 1.0.0 is not on PyPI yet, and the release tag does not exist yet. Use the canonical-v2-migration branch. The 1.3 S3 bundles are published.

Use the release process for current preparation, fit, evaluation, and publication steps.

Choose a workflow

The model, the name-count index, and the data download from three public S3 bundles: inference, dataset, and source. See Download data. Small name tuples and ORCID prefix counts ship in normal Git and the Python wheel. The source bundle holds the full corrected records, the starting point for future corrections.

Install

S2AND supports Python 3.11, 3.12, and 3.13. The 1.0.0 package requires the exactly matched s2and-rust package. Until publication, use a source checkout:

git clone --branch canonical-v2-migration https://github.com/allenai/S2AND.git
cd S2AND
uv sync

uv sync builds the native extension from s2and_rust/. uv sync --extra dev also installs the test and lint tools. See installation prerequisites for operating-system dependencies and build recovery.

Download data

Sync the bundles into s2and/data/, the default data folder. The commands use the AWS CLI and need no AWS account.

aws s3 sync --no-sign-request s3://ai2-s2-research-public/s2and/1.3/inference/ s2and/data/
# Training, evaluation, and the quick start examples also need:
aws s3 sync --no-sign-request s3://ai2-s2-research-public/s2and/1.3/dataset/ s2and/data/

inference holds the complete model and the name-count index (2.4 GB). dataset holds the benchmark Arrow data and labels (9.1 GB). The data guide states which bundles each user needs. Git ignores the downloaded files. Classic ANDData finds the index in s2and/data/ without path configuration. Arrow and production script calls use explicit paths. Historical Arrow prefixes do not satisfy the current format-1 contract.

Quick start

After you download inference and dataset, cluster every block of a released benchmark. For qian the model puts 6,542 signatures into 1,235 clusters:

uv run python scripts/tutorial_for_predicting_with_the_prod_model.py \
    --model-path s2and/data/production_model_v1.3 \
    --arrow-dataset s2and/data/benchmarks/qian --n-jobs 8

Incremental prediction links new signatures to existing author profiles. In this example, labeled j wang signatures on papers before 2015 seed 481 profiles. The block's 308 other signatures arrive as new papers; 245 of them join a profile:

uv run python scripts/tutorial_for_predicting_with_the_prod_model.py \
    --model-path s2and/data/production_model_v1.3 \
    --arrow-dataset s2and/data/benchmarks/aminer \
    --incremental-block "j wang" \
    --seed-clusters s2and/data/benchmarks/aminer/aminer_clusters.json \
    --seed-before-year 2015 --n-jobs 8

In your own code, load the complete model from an explicit path:

from s2and.arrow_inputs import ArrowDataset
from s2and.production_model import load_production_model

clusterer = load_production_model('s2and/data/production_model_v1.3', n_jobs=8)
blocks = {'j smith': ['signature_1', 'signature_2']}
with ArrowDataset.open('/path/to/your_arrow_dataset') as dataset:
    clusters, distances = clusterer.predict_from_arrow(
        blocks,
        dataset,
        total_ram_bytes=32 * 1024**3,
    )

Replace the dataset path and signature IDs with your validated Arrow dataset. blocks maps each S2 author block to its signature IDs. The tutorial's arrow_blocks function builds that mapping from signatures.arrow. The loader rejects pickles and pairwise-only stages. The inference guide describes raw inputs, embeddings, constraints, incremental calls, and bundle compatibility.

Training and evaluation

The reproduction guide runs the production training process on the published data. The production pairwise trainer reads pairwise/model_plan.json from the synced dataset bundle. It uses the dataset names that the plan declares. The production commands describe pairwise fit, validation-only EPS selection, and incremental finalization. A historical baseline is not a prerequisite for a pairwise fit. Final release acceptance remains a separate step.

The research API guide provides smaller experiments and direct metric inspection. These examples do not replace the release test policy.

Runtime and scaling

• ANDData prediction uses Python; Arrow prediction uses Rust.
• batching_threshold enables subblocks for full prediction and caps incremental query batches.
• total_ram_bytes controls the major memory allocation bounds.
• n_jobs controls concurrency; model load defaults to one job unless you specify another value.

Set thread environment variables before heavy-library imports. See subblocking, threading, and environment settings.

Development

Run the full CI checks with uv run --no-project python scripts/run_ci_locally.py.

The tests require the native extension from the development setup. See the development guide for focused checks and native build recovery.

Reproducibility

The original paper experiments use the s2and_paper branch and Python 3.7.9. Use that branch's pinned environment and experiment commands for those results. Retired historical scripts remain in Git history; see the script catalog.

The current process starts from a complete corrected snapshot. It preserves exact pair and split identities and uses versioned native models. See the release process for compatible artifacts and update rules.

Licensing

S2AND code uses the MIT License (LICENSE). S2AND data uses the Creative Commons Attribution 4.0 International License (LICENSE-DATA). Earlier data releases retain their published licenses. Some affiliation data comes from the Microsoft Academic Graph.

Citation

Cite S2AND: A Benchmark and Evaluation System for Author Name Disambiguation when you use S2AND in research.

@inproceedings{subramanian2021s2and,
  title={{S}2{AND}: {A} {B}enchmark and {E}valuation {S}ystem for {A}uthor {N}ame {D}isambiguation},
  author={Subramanian, Shivashankar and King, Daniel and Downey, Doug and Feldman, Sergey},
  year={2021},
  publisher={Association for Computing Machinery},
  address={New York, NY, USA},
  booktitle={{JCDL} '21: Proceedings of the {ACM/IEEE} Joint Conference on Digital Libraries in 2021},
  series={JCDL '21}
}

The Allen Institute for Artificial Intelligence develops S2AND.