DATASET_CARD.md

metadata

pretty_name: UserModeling (Polymarket BehaviorBench)
task_categories:
  - text-classification
  - tabular-regression
language:
  - en
license: apache-2.0
tags:
  - behavioral-modeling
  - prediction-markets
  - user-modeling
  - polymarket

UserModeling (Polymarket BehaviorBench)

Dataset description

This dataset supports wallet-level behavioral prediction on Polymarket. It asks whether models can use a wallet's past activity to predict its future position-derived stance or next trading action. It does not use market resolution as the primary prediction target.

Tasks

• Belief: predict a dominant held outcome (choice) and a position-derived normalized holding imbalance (confidence).
• Trade: predict a next-action direction (increase, decrease, or close) and an estimated USDC amount.

Input interfaces

• Direct: behavior history is supplied directly.
• Profile: history is represented by a generated wallet profile.
• Retrieval: context is augmented with records retrieved from an external support-wallet pool.

Data organization

• dataset/Belief/source/data.parquet: 1,810,609 wallet-market stance rows.
• dataset/Belief/splits/: fixed train/validation/test instances for 2,000 experiment wallets.
• dataset/Belief/support_set/: 1,669,164 rows from 29,319 non-experiment wallets.
• dataset/Trade/source/: upstream trade, price-matched, semantic, and behavior-feature tables.
• dataset/Trade/splits/: fixed structured train/validation/test trade rows.
• dataset/Trade/support_set/: 5,086,400 trades from 11,676 non-experiment wallets.
• benchmark/: model-ready Direct, Profile, and Retrieval instances in Parquet and/or compressed JSONL.

See the repository README and docs/ for schemas and lineage.

Belief label semantics

For released source rows:

[ \text{confidence} = 2\frac{\text{stance_hold_amt}}{\text{total_hold_amt}} - 1 = \frac{H_{\max}-H_{\min}}{H_{\max}+H_{\min}}. ]

This value measures the normalized imbalance of the wallet's two-sided holdings. It is not a directly elicited subjective probability, a market price, or the probability that the market resolves to the selected outcome.

Splits and retrieval support

Belief split metadata states a 70/15/15 per-wallet, time-ordered construction using stance_first_block, stance_last_block, and market_id, with rolling history and at most 30 history items in prompts.

Belief retrieval metadata documents exact event_slug matching, a strict block-time filter (support.stance_last_block < target.stance_first_block), same-wallet exclusion, and the five latest matches.

Trade support metadata states that the fixed 2,000 experiment wallets are excluded and support trades are sorted by wallet, block, and log index.

Intended uses

• Evaluating personalized behavioral prediction from prior wallet activity.
• Comparing direct-history, profile-based, and retrieval-augmented user representations.
• Studying classification and continuous behavioral targets.

Responsible use

The release contains public pseudonymous blockchain addresses and behavioral traces. It is intended for research evaluation, not identity attribution, deanonymization, surveillance, financial advice, or consequential decisions about individuals. See docs/RESPONSIBLE_AI.md for limitations, biases, sensitive-information considerations, intended uses, social impact, synthetic profile content, and provenance.

License and machine-readable metadata

BehaviorBench is released under the Apache License 2.0; see LICENSE. Croissant 1.1 core and Responsible AI metadata are provided in croissant.json.