BehaviorBench / docs /splits_and_support.md
mingzhu0527's picture
Release BehaviorBench artifact (from lyang84/BehaviorBench-camera-ready)
f9b150b verified
|
Raw History Blame Contribute Delete
1.28 kB

Splits and support pools

Belief splits

dataset/Belief/splits/split_config_v1.json records:

  • 2,000 experiment wallets;
  • seed 20260318;
  • ratios 70% train, 15% validation, 15% test;
  • ordering keys stance_first_block, stance_last_block, market_id;
  • rolling history during split construction;
  • at most 30 history items in prompts.

Counts are 98,094 train, 21,147 validation, and 22,204 test rows. Each split contains all 2,000 experiment wallets.

Trade splits

Structured Trade counts are 1,311,941 train, 114,471 validation, and 59,560 test rows, with a fixed 2,000-wallet list.

Support pools

Belief

The support manifest states:

  • all wallets except the experiment 2,000;
  • 29,319 support wallets;
  • 1,669,164 support rows;
  • deduplication by wallet, market, and first stance block.

Belief retrieval metadata specifies:

  1. exact event_slug match;
  2. support.stance_last_block < target.stance_first_block;
  3. same-wallet exclusion;
  4. five latest eligible records.

Trade

Trade support metadata states:

  • all wallets except the fixed 2,000 experiment wallets;
  • 11,676 support wallets;
  • 5,086,400 trades;
  • sorting by wallet, block, and log index.

The released benchmark rows include retrieval_method and retrieved_market_traces.