Hugging Face's logo

Datasets:
Reppo-labs
/
marketcrowd-geopolitics

Tasks:
Text Classification
Question Answering
Text Generation
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
prediction-markets
crowdsourcing
geopolitics
forecasting
reasoning
ai-evaluation
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
marketcrowd-geopolitics / README.md
Reppo-labs's picture
Reppo-labs
Upload README.md
621ed17 verified
preview code
|
raw
history blame contribute delete
9.62 kB
metadata 
license: cc-by-4.0
language:
  - en
tags:
  - prediction-markets
  - crowdsourcing
  - geopolitics
  - forecasting
  - reasoning
  - ai-evaluation
  - agents
  - uncertainty
  - calibration
  - human-feedback
  - stake-assured
  - rlhf
task_categories:
  - text-classification
  - question-answering
  - text-generation
pretty_name: MarketCrowd Geopolitics
size_categories:
  - n<1K
configs:
  - config_name: votes
    data_files:
      - split: train
        path: data/train.jsonl
  - config_name: market_summary
    data_files:
      - split: train
        path: data/market_summary.jsonl

MarketCrowd Geopolitics

The first open dataset produced via stake-assured human feedback (SAHF) — preference signals crowdsourced through capital-at-risk voting on geopolitical AI reasoning.

Overview

MarketCrowd Geopolitics contains anonymized crowd feedback votes and market-level summaries derived from a geopolitical prediction-market workflow on the Reppo protocol.

Unlike standard annotation datasets where labelers are paid per task, every signal in this dataset was produced by voters who locked $REPPO tokens as economic collateral — meaning their judgments carry capital at risk, not just attention.

This is an initial seed release covering ~6 weeks of activity across 51 geopolitical markets (March–April 2026). The underlying dataset updates continuously as new epochs settle on the Reppo protocol.

How This Data Was Produced

This dataset was generated through a Reppo Datanet — an on-chain data market where:

  1. Publishers submit geopolitical analyses as Pods (units of content tied to a market question)
  2. Voters lock $REPPO tokens to receive veREPPO voting power, then allocate that power to evaluate each Pod — positively or negatively
  3. Linear decay rewards early, independent judgment: votes cast earlier in the 48-hour epoch carry more weight than late votes, reducing herd behavior
  4. Net-vote settlement at epoch close determines which Pods enter the curated dataset (positive net vote = accepted; negative = filtered out, publisher zeroed)
  5. The resulting consensus is a synthetic oracle — no external arbiter required

This mechanism is described in full in the Reppo whitepaper.

Why stake-assured feedback differs from standard annotation

Property Paid annotation (e.g. Scale AI) Reppo SAHF
Voter incentive Payment per task Emissions + locked capital at risk
Sybil resistance Identity verification Capital requirement (splitting stake doesn't increase power)
Disagreement handling Averaged out Explicit signal via net-vote mechanism
Quality metric Inter-annotator agreement Economic Value of Feedback (EVOF)
Data staleness Static snapshot Continuously updated every 48 hours

Voting power concentration

The voting_power field is derived from locked $REPPO × lock duration (non-linear). In this release, voting power ranges from 2,471 to 3,013,340 with a median of 64,110. The top 3 wallets account for ~13% of total voting power — relatively distributed for a token-weighted system. The Reppo protocol further mitigates concentration via square-root dampening in the EVOF metric.

What Is Included

File Rows Description
data/train.jsonl 160 Row-level anonymized weighted crowd feedback votes
data/train.csv 160 CSV version of the row-level data
data/market_summary.jsonl 51 Aggregated summaries by geopolitical question/market
data/market_summary.csv 51 CSV version of the market-level summary data
DATA_DICTIONARY.md — Field-level documentation
dataset_schema.json — Machine-readable schema

Dataset Structure

data/train.jsonl — row-level votes

Each row is an anonymized feedback vote attached to a geopolitical prediction-market question.

Field Type Description
id string Stable public row ID
source_record_id string Original feedback record ID
record_type string Always crowd_feedback_vote
question string Geopolitical market/question title
category string Always geopolitics in this release
market_id string Original market/Pod ID
epoch int 48-hour epoch in which the vote was cast (higher = more recent)
feedback string Qualitative label and/or freeform comment. Format: "Label1, Label2 - optional comment"
vote string up_vote or down_vote
up_vote bool Boolean vote direction
votes int Raw on-chain vote count. Distinct from voting_power — will be 0 for many rows
voting_power int veREPPO weight — the economically meaningful signal. Derived from $REPPO locked × lock duration
created_at string Feedback creation timestamp
updated_at string Feedback update timestamp
deleted bool Source deletion flag
version int Source version field
split string Dataset split

data/market_summary.jsonl — market-level aggregates

One row per geopolitical question, summarizing all votes for that market.

Field Type Description
id string Stable public summary ID
record_type string Always market_feedback_summary
question string Geopolitical market/question title
category string Topic category
market_id string Original market/Pod ID
num_feedback_votes int Total number of feedback votes
up_votes int Count of positive votes
down_votes int Count of negative votes
up_vote_share float Fraction of votes that were positive
total_voting_power int Sum of voting power across all votes
median_voting_power float Median voting power per vote
top_feedback_tags list Most common feedback labels
first_seen_at string Earliest feedback timestamp
last_seen_at string Latest feedback timestamp
split string Dataset split

Example Row 

{
  "id": "mcg_vote_000001",
  "source_record_id": "cmndwvyey0001kz04mn8e5fmp",
  "record_type": "crowd_feedback_vote",
  "question": "Hormuz Updates",
  "category": "geopolitics",
  "market_id": "cmn9plf2k0001kz040cuuizdq",
  "epoch": 64,
  "feedback": "High quality -",
  "vote": "up_vote",
  "up_vote": true,
  "votes": 0,
  "voting_power": 1592667,
  "created_at": "2026-03-31 01:01:42.778",
  "updated_at": "2026-03-31 01:01:42.778",
  "deleted": false,
  "version": 1,
  "split": "train"
}

Suggested Uses

This dataset is best treated as a feedback-signal and question-quality dataset. It should not be used as a source of resolved geopolitical ground truth.

Recommended training setup:

Component Field
Input question + feedback
Label up_vote or vote
Sample weight voting_power

Agent training use cases:

Agent task Input Signal
Question quality scoring Market question up_vote_share, voting_power, feedback labels
Market triage Question + feedback Aggregated market summary
Feedback classification Feedback text Vote direction or feedback category
Signal summarization Question + row-level votes Market-level summary
Forecasting workflow support Proposed market Crowd-perceived usefulness and clarity

Other applications: text classification, weak supervision, market ranking, crowd-signal summarization, geopolitical forecasting prompt evaluation, prediction-market-based data research.

Limitations

  • Size: 160 votes across 51 markets from a 6-week window. This is a seed release — not yet suitable as a standalone benchmark.
  • Single domain: All records are category: geopolitics from one Datanet. Cross-domain generalization should not be assumed.
  • Feedback field: Mixes structured tags with freeform comments. Preprocessing recommended before classification tasks.
  • No resolved outcomes: This is a preference/feedback dataset, not a ground-truth geopolitical dataset. Event resolution is not included.
  • votes vs voting_power: The votes field is 0 for most rows — it reflects a raw on-chain count that was not recorded for all votes. Use voting_power as the primary signal.

Privacy

The source data contained direct voter and subnet identifiers. These were removed before this public release:

  • voter_id — removed
  • private_subnet_id — removed

About Reppo

Reppo is a protocol for tokenized, continuously curated data production for reinforcement learning. Datanets are on-chain data markets where publishers contribute content and veREPPO holders provide stake-assured quality assessments. Datasets update every 48 hours and are available for subscription via repo.exchange.

Citation 

@dataset{reppo_marketcrowd_geopolitics_2026,
  author    = {Reppo Labs},
  title     = {MarketCrowd Geopolitics: Stake-Assured Human Feedback on Geopolitical AI Reasoning},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Reppo-labs/marketcrowd-geopolitics}
}

License

CC-BY-4.0 — attribution to Reppo Labs and the MarketCrowd Geopolitics dataset required.