This repository provides access to the S2AND dataset and S2AND reference model described in the paper S2AND: A Benchmark and Evaluation System for Author Name Disambiguation by Shivashankar Subramanian, Daniel King, Doug Downey, Sergey Feldman.

The reference model is live on semanticscholar.org, and the trained model is available now as part of the data download (see below).

Clone the repo.

Install uv using the official guide:

• uv installation docs

Install Rust (needed to build the native extension from source):

• Rust installation docs

If you are building the Rust extension, install OS prerequisites:

# Ubuntu / Debian / WSL2
sudo apt-get update
sudo apt-get install -y build-essential pkg-config libgomp1

# Windows (one-time)
# Install Visual Studio Build Tools with the "Desktop development with C++" workload.

Verify toolchain availability:

uv --version
rustc --version
cargo --version

WSL notes:

• Some Ubuntu images do not provide a python alias by default; use python3 for system Python commands.
• On PEP 668-managed systems, python3 -m pip install --user ... may fail with externally-managed-environment; use one of the official uv install methods above.

1. From repo root:

# create the project venv (uv defaults to .venv if you don't give a name)
# use Python 3.11.x (fasttext doesn't support 3.12+ here)
uv venv --python 3.11.13

2. Activate the venv (choose one):

# macOS / Linux (bash / zsh)
source .venv/bin/activate

# Windows PowerShell
. .venv\Scripts\Activate.ps1

# Windows CMD
.venv\Scripts\activate.bat

3. Runtime install (end users, pick one):

# default runtime is Python (`auto` resolves to Python).
uv pip install s2and
# optional: Rust-enabled runtime when extension wheels are available.
uv pip install "s2and[rust]"

4. Developer install (repo checkout):

# prefer uv --active so uv uses your activated environment
uv sync --active --extra dev

5. (Recommended) Build/install the Rust extension into the active venv:

# requires Rust toolchain on PATH (rustc/cargo)
# if Rust was just installed via rustup in this shell:
# source "$HOME/.cargo/env"
uv run --active --no-project maturin develop -m s2and_rust/Cargo.toml

Notes:

• This installs the native module into site-packages so imports use the compiled extension.
• If you don't want an editable install, you can uv pip install . instead of uv sync, then run the maturin develop step above.
• Once wheels are published, you can install the native extension via extras: uv pip install "s2and[rust]".
• On WSL with repo paths mounted from Windows (for example, /mnt/c/...), uv may warn about failed hardlinks. To suppress this and avoid repeated warnings, set UV_LINK_MODE=copy before uv sync / uv pip install.

• Index (start here): docs/README.md
• Next steps: docs/work_plan.md
• Backlog: docs/work_plan.md (Backlog section)

To run the tests, use the following command:

uv run --no-project pytest tests/

To run the entire CI suite mimicking the GH Actions, use the following command:

uv run python scripts/run_ci_locally.py

scripts/run_ci_locally.py mirrors .github/workflows/main.yaml by running:

• lint job (ruff check + ruff format --check)
• typecheck-and-test matrix lanes (py-only, then rust-enabled)
• Rust parity guardrail tests in the rust-enabled lane

By default, local ty checks use --python-version 3.11 --python-platform linux to match GitHub Linux runners. To override platform emulation locally, set S2AND_CI_TY_PLATFORM (for example, windows).

To run CI checks locally without Rust extension compilation (faster iteration), run:

uv sync --active --extra dev --frozen
uv run --active --no-project ruff format --check s2and scripts/*.py
uv run --active --no-project ty check s2and --ignore unresolved-import --ignore unused-type-ignore-comment --ignore possibly-missing-attribute --ignore unresolved-global
uv run --active --no-project ty check scripts/*.py --ignore unresolved-import --ignore unused-type-ignore-comment --ignore possibly-missing-attribute --ignore unresolved-global --ignore unresolved-reference --ignore unresolved-attribute
# macOS/Linux:
PYTHONPATH=. uv run --active --no-project pytest tests/ --cov=s2and --cov-report=term-missing --cov-fail-under=40
# Windows PowerShell:
$env:PYTHONPATH='.'; uv run --active --no-project pytest tests/ --cov=s2and --cov-report=term-missing --cov-fail-under=40

Versioning is centralized in the VERSION file (single source of truth). When you update it, we sync the Python/Rust manifests and regenerate lockfiles.

One-time setup for hooks (recommended):

git config core.hooksPath .githooks

Workflow:

# 1) edit VERSION
echo 0.40.0 > VERSION

# 2) sync manifests
uv run python scripts/sync_version.py

# 3) regenerate lockfiles
uv sync --extra dev
uv run --active --no-project cargo generate-lockfile --manifest-path s2and_rust/Cargo.toml

Notes:

• The pre-commit hook only runs when VERSION is staged and will auto-sync + regenerate lockfiles if needed.
• uv.lock and s2and_rust/Cargo.lock are generated files and will contain the version after syncing.

When running scripts from the repo, prefer uv run --no-project so the installed packages (including the Rust extension) resolve from site-packages. Avoid setting PYTHONPATH to the repo root, which can shadow the compiled module.

Quickstart with bundled fixture (tests/qian, no large data download):

uv run --no-project python scripts/tutorial_for_predicting_with_the_prod_model.py \
  --use-rust 1 \
  --dataset qian \
  --data-root tests \
  --load-name-counts 0

Run the same tutorial on data/s2and_mini (downloaded artifacts):

uv run --no-project python scripts/tutorial_for_predicting_with_the_prod_model.py --use-rust 1 --dataset qian

Profiling (Rust inference):

S2AND_BACKEND=rust uv run --no-project python scripts/rust_suite.py prod-inference \
  --dataset-name qian \
  --data-root tests \
  --n-jobs 4

Benchmark baseline ownership:

• Active Rust runtime gate baselines and promotion rules: docs/rust/baselines.md

S2AND backend selection is controlled by S2AND_BACKEND:

• auto (default) — uses Rust when available and capable, otherwise Python
• rust — strict Rust mode; fails fast on Rust-stage errors
• python — Python-only path; zero Rust calls

For the full list of environment variables, see docs/environment.md.

Install contract:

• uv pip install s2and: Python-only runtime.
• uv pip install "s2and[rust]": Rust-enabled runtime.
• Full runtime contract: docs/rust/runtime.md.

Notes:

• Rust batch mode uses Rayon internally for parallelism; Python process pools are not used.
• When Rust is enabled, signature n-gram Counters may be deferred and computed natively during Rust featurizer construction.
• If a Python code path needs eager n-gram Counters, call ANDData.materialize_signature_ngrams_python().

• Default: use_cache=False (no caching).
• use_cache=True: enables Python pair-feature cache and Rust featurizer cache.
• Cache root: S2AND_CACHE env var (defaults to ~/.s2and).

Pre-warm once at server start:

from s2and.feature_port import warm_rust_featurizer
warm_rust_featurizer(dataset, use_cache=True)

For processing massive blocks (hundreds of thousands of signatures), use the Rust backend with subblocking to keep memory bounded. This is the recommended production setup.

Use predict() with batching_threshold to automatically split large blocks into manageable subblocks:

import os

# 1. Force Rust backend (set before importing s2and modules)
os.environ["S2AND_BACKEND"] = "rust"

from s2and.data import ANDData
from s2and.feature_port import warm_rust_featurizer
from s2and.serialization import load_pickle_with_verified_label_encoder_compat

# 2. Load the production model
clusterer = load_pickle_with_verified_label_encoder_compat(
    "data/production_model_v1.2.pickle"
)["clusterer"]
clusterer.use_cache = False  # disable caching for one-shot inference
clusterer.n_jobs = 8

# 3. Load your dataset in inference mode
dataset = ANDData(
    signatures="path/to/signatures.json",
    papers="path/to/papers.json",
    specter_embeddings="path/to/specter.pickle",
    mode="inference",
    block_type="s2",
    n_jobs=8,
    name="my_dataset",
)

# 4. (Optional) Pre-warm Rust featurizer to reduce cold-start latency
warm_rust_featurizer(dataset, use_cache=False)

# 5. Predict clusters with subblocking for large blocks
pred_clusters, _ = clusterer.predict(
    dataset.get_blocks(),
    dataset,
    batching_threshold=5000,  # blocks larger than this are split into subblocks
    desired_memory_use=5000 * 5000,  # memory budget in signature-pairs (25M pairs here)
)

# pred_clusters is a dict mapping signature_id -> list of cluster member signature_ids
print(f"Total clusters: {len(pred_clusters)}")

Key parameters for predict():

• batching_threshold: blocks larger than this are split via make_subblocks() before clustering
• desired_memory_use: memory budget in signature-pair units; controls chunk sizing for subblocked incremental paths (default: batching_threshold²)

Use predict_incremental() when you have existing clusters (cluster_seeds) and want to assign new signatures without reclustering everything:

# Load dataset with existing cluster seeds
dataset = ANDData(
    signatures="path/to/signatures.json",
    papers="path/to/papers.json",
    specter_embeddings="path/to/specter.pickle",
    mode="inference",
    block_type="s2",
    n_jobs=8,
    name="my_dataset",
    cluster_seeds={
        "require": {
            "block_key": {("sig1", "sig2"): 1.0, ...},  # pairs that must cluster together
        },
        "disallow": {
            "block_key": {("sig3", "sig4"), ...},  # pairs that must NOT cluster together
        },
    },
)

# Run incremental prediction on one block
blocks = dataset.get_blocks()
block_key = "j smith"  # target block
block_signatures = blocks[block_key]

result = clusterer.predict_incremental(
    block_signatures,
    dataset,
    batching_threshold=5000,       # subblock size cap for phase-split mode
    total_ram_bytes=32 * 1024**3,  # explicit RAM budget (32 GB)
)

clusters = result["clusters"]
phase_b_mode = result.get("phase_b_mode", "N/A")

# phase_b_mode indicates how phase-split handled memory:
# - "exact": ran Phase B globally (monolithic-equivalent behavior)
# - "subblock_local": ran Phase B per-subblock (memory-bounded approximation)
print(f"Clusters: {len(set(clusters.values()))}, mode={phase_b_mode}")

For detailed subblocking behavior, see docs/subclustering.md.

To obtain the S2AND dataset, run the following command after the package is installed (from inside the S2AND directory): [Expected download size is: 50.4 GiB]

aws s3 sync --no-sign-request s3://ai2-s2-research-public/s2and-release data/

Note that this software package comes with tools specifically designed to access and model the dataset.

Modify the config file at data/path_config.json. This file should look like this

{
    "main_data_dir": "absolute path to wherever you downloaded the data to",
    "internal_data_dir": "ignore this one unless you work at AI2"
}

As the dummy file says, main_data_dir should be set to the location of wherever you downloaded the data to, and internal_data_dir can be ignored, as it is used for some scripts that rely on unreleased data, internal to Semantic Scholar.

Once you have downloaded the datasets, you can go ahead and load up one of them:

from os.path import join
from s2and.data import ANDData

dataset_name = "pubmed"
parent_dir = f"data/{dataset_name}"
dataset = ANDData(
    signatures=join(parent_dir, f"{dataset_name}_signatures.json"),
    papers=join(parent_dir, f"{dataset_name}_papers.json"),
    mode="train",
    specter_embeddings=join(parent_dir, f"{dataset_name}_specter.pickle"),
    clusters=join(parent_dir, f"{dataset_name}_clusters.json"),
    block_type="s2",
    train_pairs_size=100000,
    val_pairs_size=10000,
    test_pairs_size=10000,
    name=dataset_name,
    n_jobs=8,
)

This may take a few minutes - there is a lot of text pre-processing to do.

The first step in the S2AND pipeline is to specify a featurizer and then train a binary classifier that tries to guess whether two signatures are referring to the same person.

We'll do hyperparameter selection with the validation set and then get the test area under ROC curve.

Here's how to do all that:

from s2and.model import PairwiseModeler
from s2and.featurizer import FeaturizationInfo, featurize
from s2and.eval import cluster_eval, pairwise_eval

featurization_info = FeaturizationInfo()
# the cache will make it faster to train multiple times - it stores the features on disk for you
train, val, test = featurize(dataset, featurization_info, n_jobs=8, use_cache=True)
X_train, y_train = train
X_val, y_val = val
X_test, y_test = test

# calibration fits isotonic regression after the binary classifier is fit
# monotone constraints help the LightGBM classifier behave sensibly
pairwise_model = PairwiseModeler(
    n_iter=25, calibrate=True, monotone_constraints=featurization_info.lightgbm_monotone_constraints
)
# this does hyperparameter selection, which is why we need to pass in the validation set.
pairwise_model.fit(X_train, y_train, X_val, y_val)

# this will also dump a lot of useful plots (ROC, PR, SHAP) to the figs_path
pairwise_metrics = pairwise_eval(X_test, y_test, pairwise_model.classifier, figs_path='figs/', title='example')
print(pairwise_metrics)

The second stage in the S2AND pipeline is to tune hyperparameters for the clusterer on the validation data and then evaluate the full clustering pipeline on the test blocks.

We use agglomerative clustering as implemented in fastcluster with average linkage. There is only one hyperparameter to tune.

from s2and.model import Clusterer, FastCluster
from hyperopt import hp

clusterer = Clusterer(
    featurization_info,
    pairwise_model,
    cluster_model=FastCluster(linkage="average"),
    search_space={"eps": hp.uniform("eps", 0, 1)},
    n_iter=25,
    n_jobs=8,
)
clusterer.fit(dataset)

# the metrics_per_signature are there so we can break out the facets if needed
metrics, metrics_per_signature = cluster_eval(dataset, clusterer)
print(metrics)

For a fuller example, please see the transfer script: scripts/transfer_experiment_seed_paper.py.

Assuming you have a clusterer already fit, you can dump the model to disk like so

import pickle

with open("saved_model.pkl", "wb") as _pkl_file:
    pickle.dump(clusterer, _pkl_file)

You can then reload it, load a new dataset, and run prediction

import pickle

with open("saved_model.pkl", "rb") as _pkl_file:
    clusterer = pickle.load(_pkl_file)

anddata = ANDData(
    signatures=signatures,
    papers=papers,
    specter_embeddings=paper_embeddings,
    name="your_name_here",
    mode="inference",
    block_type="s2",
)
pred_clusters, pred_distance_matrices = clusterer.predict(anddata.get_blocks(), anddata)
# pred_distance_matrices can be None when using memory-optimized fused clustering

We provide trained production models in the S3 bucket along with the datasets:

• production_model_v1.2.pickle — current production model (used on Semantic Scholar website and API). It expects SPECTER2 [PRX] embeddings for papers. It does not use reference features.
• production_model_v1.1.pickle — previous production version. It expects SPECTER1 embeddings for papers. It does not use reference features.
• production_model_v1.0.pickle — original version (deprecated). It expects SPECTER1 embeddings for papers. It does use reference features.

To see an example of using it, please see the script scripts/tutorial_for_predicting_with_the_prod_model.py. You can also use it on your own data, as long as it is formatted the same way as the S2AND data.

Bundled quick example:

uv run --no-project python scripts/tutorial_for_predicting_with_the_prod_model.py \
  --use-rust 1 \
  --dataset qian \
  --data-root tests \
  --load-name-counts 0

Please note that the production models still use SPECTER1, and these embeddings are still available via the S2 API.

S2AND currently supports two runtime semantics for the name-count feature key used by last_first_initial_count_min:

• legacy_full_first_token: key is <last> <first_token> (historical behavior).
• initial_char: key is <last> <first[0]> (current intended semantics).

Model compatibility rules:

• production_model_v1.1.pickle and production_model_v1.2.pickle were trained with legacy_full_first_token.
• In ANDData(..., mode="inference"), prediction automatically applies the semantics expected by the loaded model via clusterer.feature_contract["name_counts_last_first_initial_semantics"] (with featurizer_version fallback for older artifacts).
• Do not mix model artifacts and feature semantics without retraining, because this changes model inputs and can materially change clustering output.

There is a also a predict_incremental function on the Clusterer, that allows prediction for just a small set of new signatures. When instantiating ANDData, you can pass in cluster_seeds, which will be used instead of model predictions for those signatures. If you call predict_incremental, the full distance matrix will not be created, and the new signatures will simply be assigned to the cluster they have the lowest average distance to, as long as it is below the model's eps, or separately reclustered with the other unassigned signatures, if not within eps of any existing cluster.

For very large incremental blocks, phase-split mode is used automatically when subblocking is active (i.e., when batching_threshold is set and the block exceeds it). Phase-split subblocks Phase A and then:

• runs Phase B globally when it fits budget (phase_b_mode="exact"),
• auto-falls back to subblock-local B/C/D when over budget (phase_b_mode="subblock_local").

predict_incremental returns a payload with:

• clusters
• phase_b_mode
• phase_b_budget_bytes
• phase_b_required_bytes

RAM policy:

• Preferred: pass total_ram_bytes=<int> directly to predict_incremental.
• If omitted, runtime auto-detects RAM (cgroup first, then host probes) and applies a 0.8 safety factor before deriving budgets.

The experiments in the paper were run with the python (3.7.9) package versions in paper_experiments_env.txt, in the branch s2and_paper.

To install, run:

git checkout s2and_paper
pip install pip==21.0.0
pip install -r paper_experiments_env.txt --use-feature=fast-deps --use-deprecated=legacy-resolver

Then, rerunning scripts/paper_experiments.sh on the branch s2and_paper should produce the same numbers as in the paper (we will update here if this becomes not true).

Our trained, released models are in the s3 folder referenced above, and are called production_model.pickle (very close to what is running on the Semantic Scholar website, except the production model doesn't compute the reference features) and full_union_seed_*.pickle (models trained during benchmark experiments). They can be loaded the same way as in the section above called "How to use S2AND for predicting with a saved model", except that the pickled object is a dictionary, with a clusterer key. Important: these pickles will only run on the branch s2and_paper and not on main.

use_cache defaults to False in the current codebase. Enable it explicitly when you want cached reruns, and disable it when validating feature changes or running one-shot experiments.

The code in this repo is released under the Apache 2.0 license. The dataset is released under ODC-BY (included in S3 bucket with the data). We would also like to acknowledge that some of the affiliations data comes directly from the Microsoft Academic Graph (https://aka.ms/msracad).

If you use S2AND in your research, please cite S2AND: A Benchmark and Evaluation System for Author Name Disambiguation.

@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}
}

S2AND is an open-source project developed by the Allen Institute for Artificial Intelligence (AI2). AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.