Forecast-Dojo

Replayable forecasting tasks for training and evaluating LLM agents.

Resolved Polymarket events paired with fixed historical forecast dates, outcome labels, and contemporaneous market probabilities.

Split	Events	Event–date tasks	Binary / multi-option
Train	1,338	5,325	602 / 736
Eval	230	797	40 / 190

The splits contain disjoint events. Each event has at least three forecast dates, all at least two days before its recorded close.

Format

Each JSONL row is one event with an ordered forecast schedule.

Field	Description
event_id, slug	Event identifiers
belief_kind	binary or multi_neg_risk
prompt, metadata.question	Question text, resolution criteria, and offered outcomes
label	Resolved outcome
metadata.start_date	Retrieval-window start
metadata.forecast_dates	Ordered forecast dates
metadata.crowd_probs	Market probability of the resolved outcome at each date
metadata.crowd_distributions	Market outcome distributions aligned with the dates

Unavailable market references are null; missing multi-option prices contribute zero to the market vectors. The resolved label and market references are for scoring, not agent input.

Load

Read the original JSONL to preserve each event's outcome dictionary:

import json
from huggingface_hub import hf_hub_download

split = "eval"  # "train" or "eval"
path = hf_hub_download(
    repo_id="liqinye/Forecast-Dojo",
    filename=f"data/{split}.jsonl",
    repo_type="dataset",
)
with open(path, encoding="utf-8") as f:
    events = [json.loads(line) for line in f if line.strip()]

Source: Polymarket. This release contains event data; the news corpus, retrieval indexes, and model checkpoints are not included.