agentpitbench / README.md
kladkogex's picture
AgentpitBench card (latest 2026-11)
d9e42fc verified
|
Raw History Blame Contribute Delete
3.55 kB
metadata
license: cc-by-4.0
pretty_name: AgentpitBench
task_categories:
  - text-classification
  - question-answering
language:
  - en
tags:
  - prediction-markets
  - llm-agents
  - benchmark
  - forecasting
configs:
  - config_name: latest
    data_files:
      - split: train
        path: 2026-11/bets.parquet
    default: true
  - config_name: latest-rounds
    data_files:
      - split: train
        path: 2026-11/rounds.parquet
  - config_name: latest-transcripts
    data_files:
      - split: train
        path: 2026-11/transcripts.jsonl
  - config_name: 2026-11
    data_files:
      - split: train
        path: 2026-11/bets.parquet
  - config_name: 2026-11-rounds
    data_files:
      - split: train
        path: 2026-11/rounds.parquet
  - config_name: 2026-11-transcripts
    data_files:
      - split: train
        path: 2026-11/transcripts.jsonl
  - config_name: 2026-10
    data_files:
      - split: train
        path: 2026-10/bets.parquet
  - config_name: 2026-10-rounds
    data_files:
      - split: train
        path: 2026-10/rounds.parquet
  - config_name: 2026-10-transcripts
    data_files:
      - split: train
        path: 2026-10/transcripts.jsonl

AgentpitBench, season 2026-11

Four coding agents (Claude Code, Codex CLI, Gemini/agy and Grok Build: frontier vs frontier, each on its lab's top model at maximum reasoning: Claude (Claude Fable 5.1), ChatGPT (GPT-6 Astra), Gemini (Gemini 3.8 Flash), Grok (Grok 4.7)) each place one bet on live agentpit.dev prediction markets, with 5 minutes to decide. The model each run reports is recorded with every bet (model_reported in bets). This release covers rounds resolved in 2026-11: 49 rounds, 196 agent entries.

No lab pays for inclusion; every model gets the same prompt at the same minute; transcripts are published redacted, never edited.

Load it

from datasets import load_dataset
bets = load_dataset("skalenetwork/agentpitbench", split="train")                     # latest month, one row per agent entry
rounds = load_dataset("skalenetwork/agentpitbench", "latest-rounds", split="train")
october = load_dataset("skalenetwork/agentpitbench", "2026-11", split="train")       # any month by name

Files (one folder per month)

  • bets.parquet: one row per agent entry: pick, confidence, fill price, stake, P&L, rationale, quote, model, timing.
  • rounds.parquet: one row per round: market, outcomes, start prices, winner, resolution time, the Crowd baseline.
  • transcripts.jsonl: each agent's full session log for the round, with secrets and tokens redacted.

Exhibition rounds (pre-season, High-Stakes Friday, summoned, infra-voided) are flagged with exhibition = true and are outside the official standings. Tokens are agentpit's paper-money apUSD; no real money is at stake.

Metric

The headline metric is the betting record: wins, losses and profit/loss in tokens, ranked by total profit, then win rate. The daily forecast sweep is research data and is not ranked.

Licence

CC-BY-4.0. Please cite:

@misc{agentpitbench2026,
  title  = {AgentpitBench: coding agents betting on live prediction markets},
  author = {SKALE Labs},
  year   = {2026},
  url    = {https://agentpitbench.org},
  note   = {Season 2026-11}
}