Datasets:

Salesforce
/

BehaviorBench

BehaviorBench is intended for research on whether models can predict the next observable action of a pseudonymous Polymarket wallet from its prior public behavior. It supports comparisons of direct-history, generated-profile, and retrieval-augmented representations.

The benchmark measures behavioral predictability, not a person's internal mental state. Belief confidence is a normalized position-imbalance score derived from holdings; it is not a directly elicited subjective probability. Trade amount is an estimated USDC value derived from token amount and mark price.

Recommended uses include benchmark evaluation, behavioral-sequence modeling, representation studies, and robustness analysis. It is not intended for identity attribution, individual financial advice, credit or employment decisions, law enforcement, or automated targeting of wallet owners.

Data sources and collection

The data are derived from public Polymarket and Polygon on-chain records. Released source tables cover Polygon block heights approximately 30,000,825 through 80,000,000. The repository contains processed source tables rather than a complete raw blockchain archive.

No survey respondents or human annotators supplied the task labels. Labels are deterministic transformations of observed holdings and transactions. Generated wallet profiles are synthetic summaries produced by language models and are identified by generator-model fields.

Preprocessing and labeling

Belief records aggregate wallet-market holdings into a stance outcome and normalized position imbalance.
Trade records order transactions by block and log index, reconstruct position changes, and derive increase/decrease/close actions.
Experiment wallets are divided into fixed, per-wallet temporal train/validation/test splits.
Retrieval support pools exclude the fixed experiment wallets.
Model-ready files convert structured records into prompts and structured targets.

Detailed schemas and lineage are documented in docs/data_construction.md, docs/schema_belief.md, docs/schema_trade.md, docs/splits_and_support.md, and docs/interfaces.md.

Personal or sensitive information

The release contains public pseudonymous blockchain wallet addresses and behavioral traces. Market participation can reveal or suggest financial interests, risk preferences, political views, sports interests, or other sensitive inferences. The dataset does not contain verified real-world identities, demographic attributes, private account data, or direct contact information.

Users must not attempt to deanonymize wallets, link addresses to named individuals, or use the data to make consequential decisions about people.

Biases

Only active Polymarket wallets satisfying task-specific activity filters are represented.
Wallet activity is not representative of the general population.
High-activity users, automated strategies, market makers, hedgers, and arbitrageurs may be overrepresented.
Market topics and participation rates are time-dependent and platform-specific.
Position-derived labels may reflect liquidity, hedging, budget, and execution constraints rather than only directional belief.
Generated profiles inherit biases and omissions from their generator models.
Retrieval quality depends on support-pool coverage and matching rules.

Limitations

Observable behavior is an imperfect proxy for preferences or beliefs.
Belief confidence is highly concentrated at 0 and 1.
Exact position ties do not identify a unique directional stance from holdings alone.
Trade values use estimated mark-price conversion rather than complete execution-cost reconstruction.
Public source tables do not reproduce every upstream crawler and indexing step.
Results may not generalize to other platforms, time periods, financial products, or populations.
Benchmark performance must not be interpreted as evidence that a model understands a person's private intentions.

Social impact

Potential benefits include improved evaluation of personalized systems, clearer measurement of behavioral-history representations, and reproducible research on real-world decision traces. Potential harms include surveillance, profiling, deanonymization, manipulation, and overclaiming psychological inference from financial behavior. Mitigations include using only public records, documenting proxy-label semantics, excluding real-world identity attributes, and explicitly prohibiting identity attribution and consequential use.

Synthetic data

The underlying holdings and transaction records are observational rather than synthetic. Profile interface fields contain model-generated summaries; model rationales are generated outputs rather than ground-truth annotations. Generator identity is retained where available.

Maintenance

The canonical release target is https://huggingface.co/datasets/Salesforce/BehaviorBench. Material corrections should be versioned and recorded in CHANGELOG.md; checksums and Croissant metadata should be regenerated for each release.