README.md

metadata

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

BehaviorBench: Polymarket Behavioral Modeling Data

This repository contains data for modeling wallet-level behavior on Polymarket. The prediction target is a wallet's behavior—not the eventual resolution of a market.

Two tasks are provided:

• Belief: predict a wallet's dominant outcome holding and a position-derived confidence score.
• Trade: predict the direction and estimated size of a wallet's next trade.

Each task is released at two levels:

• dataset/: structured source tables, fixed experimental splits, and retrieval support pools.
• benchmark/: model-ready instances for Direct, Profile, and Retrieval interfaces.

The Belief confidence field is a normalized position imbalance derived from holdings. See Belief schema and labels.

Repository layout

dataset/
  Belief/
    source/       # wallet-market holding labels
    splits/       # fixed benchmark splits for 2,000 experiment wallets
    support_set/  # external wallets used by retrieval
  Trade/
    source/       # upstream trade, price, and behavior-feature tables
    splits/       # fixed benchmark splits for 2,000 experiment wallets
    support_set/  # external wallet trade histories used by retrieval
benchmark/
  Belief/{direct,profile,retrieval}/
  Trade/{direct,profile,retrieval}/
metadata/
docs/

This camera-ready artifact contains the Polymarket BehaviorBench data only; the separate Manifold extension is not part of this release.

Dataset summary

Belief

A Belief source row represents one wallet's position-derived stance in one market. The benchmark asks a model to predict:

{"choice": "an outcome", "confidence": 0.0, "rationale": "..."}

The fixed split contains:

Split	Rows	Wallets
train	98,094	2,000
validation	21,147	2,000
test	22,204	2,000

The source table has 1,810,609 rows. The retrieval support set excludes the 2,000 experiment wallets and contains 1,669,164 rows from 29,319 wallets.

Trade

A Trade benchmark row represents a target transaction together with preceding wallet history and decision context. The target contains:

{"direction": "increase | decrease | close", "amount_usdc_est": 0.0}

The structured split contains:

Split	Rows
train	1,311,941
validation	114,471
test	59,560

The model-ready Trade benchmark files use stable sample_id values for alignment across interfaces. The retrieval support set excludes the fixed experiment wallets and contains 5,086,400 trades from 11,676 wallets.

Benchmark interfaces

• Direct: supplies wallet behavior history directly.
• Profile: supplies a generated wallet profile (and, depending on task, limited current context).
• Retrieval: augments the instance with records retrieved from a support pool that excludes experiment wallets.

See Benchmark interfaces for exact released columns and documented retrieval behavior.

Loading examples

import pandas as pd

belief = pd.read_parquet("dataset/Belief/splits/test.parquet")
trade = pd.read_parquet("benchmark/Trade/direct/test.parquet")

print(belief[["sample_id", "label_choice", "label_confidence"]].head())
print(trade[["sample_id", "input_text", "target"]].head())

To stream compressed JSONL:

import gzip, json

with gzip.open("benchmark/Belief/direct/test.jsonl.gz", "rt") as f:
    first = json.loads(next(f))

Documentation

• Data construction and lineage
• Belief schema and label semantics
• Trade schema and label semantics
• Splits and support pools
• Direct, Profile, and Retrieval interfaces

NeurIPS 2026 artifact metadata

• License: Apache-2.0
• Machine-readable metadata: croissant.json
• Responsible AI documentation: docs/RESPONSIBLE_AI.md
• Artifact checklist: docs/NEURIPS_2026_ARTIFACT_CHECKLIST.md
• Version history: CHANGELOG.md

The canonical release target is https://huggingface.co/datasets/Salesforce/BehaviorBench. After upload, regenerate or download Hugging Face Croissant metadata, merge the RAI fields, and validate the final public URLs with the NeurIPS Croissant validator before OpenReview submission.

For the controlled handoff from the temporary repository to Salesforce/BehaviorBench, follow docs/SALESFORCE_MIGRATION.md and use scripts/prepare_salesforce_release.py. This prevents accidental inclusion of manifold/ and checks for unresolved Git LFS pointers.