fireworks-ai/kalshi-decision-v1

Kalshi Decision Tuning v1

Cleaned, labeled examples from settled Kalshi prediction markets, used in the Fireworks decision-tuning cookbook.

Files and splits

File	Events (JSONL rows)	Questions	Purpose
train.jsonl	6,461	20,051	Training
dev.jsonl	2,000	6,232	Calibration; fit per-question-type temperatures
test.jsonl	8,765	33,712	Consolidated held-out evaluation
unseen_family.jsonl	765	3,567	Unseen-topic slice of the test set

unseen_family.jsonl is already included in test.jsonl: it is the exact final 765 rows. Do not append it to the test set or count it as an independent test set. Its 113 topic families were excluded from training and calibration. There are 17,226 unique events across training, calibration, and the consolidated test set.

These are byte-for-byte copies of the original public Fireworks uploads. Their row counts and SHA-256 hashes match the pins in the cookbook's prepare_data.py. manifest.json and SHA256SUMS record these checksums.

Download

Download the original JSONL files to retain the decision-training row format:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="fireworks-ai/kalshi-decision-v1",
    repo_type="dataset",
    filename="train.jsonl",
    token=False,
)

with open(path, encoding="utf-8") as handle:
    first_event = json.loads(next(handle))

For reproducible experiments, set revision to a dataset commit SHA and check the downloaded file against SHA256SUMS. The cookbook downloader verifies both SHA-256 and row counts.

Row format

One JSONL row represents one market event and contains:

• state: shared event context, including the market description and rules.
• questions: a mapping of question IDs to instructions, type, label, and weight. Multiple-choice questions also include their answer criteria.

Mutually exclusive outcomes use one choice question with a string label. Other contracts use up to eight noul (yes/no) questions with Boolean labels. The examples preserve this nested format for use with the Fireworks decision training recipe.

Training, calibration, and evaluation are separate. The model inputs use open-time market information; settled outcomes provide the labels. Data preparation removed near-duplicate examples and later changes that could reveal the outcome. See the cookbook for the full split and cleaning methodology.

Source and resources

• Source data: Kalshi public market API.
• Reproducible example: Fireworks Kalshi decision tuning.
• Original Fireworks dataset IDs: accounts/fireworks/datasets/kalshi-decision-v1-train, accounts/fireworks/datasets/kalshi-decision-v1-dev, accounts/fireworks/datasets/kalshi-decision-v1-test, and accounts/fireworks/datasets/kalshi-decision-v1-unseen-family.

These examples support historical evaluation of decision models. They contain settled outcomes and should be kept separate from live forecasting inputs.