Datasets:

greenday123
/

StakeBench

license: cc-by-4.0
language:
  - en
pretty_name: StakeBench
tags:
  - prediction-markets
  - finance
  - benchmark
  - llm-evaluation
  - polymarket
  - manifold
  - revealed-preference
task_categories:
  - text-classification
size_categories:
  - 100K<n<1M
language_creators:
  - found
annotations_creators:
  - found
source_datasets:
  - original
configs:
  - config_name: polymarket_comments
    default: true
    data_files: polymarket/comments_clean.csv
  - config_name: manifold_comments
    data_files: manifold/comments_clean.csv
  - config_name: polymarket_positions
    data_files: polymarket/positions_clean.csv
  - config_name: manifold_positions
    data_files: manifold/positions_clean.csv
  - config_name: polymarket_markets
    data_files: polymarket/market_attrs.csv
  - config_name: manifold_markets
    data_files: manifold/market_attrs.csv
  - config_name: polymarket_topics
    data_files: polymarket/topics_summary.csv
  - config_name: manifold_markets_summary
    data_files: manifold/markets_summary.csv
  - config_name: polymarket_prices
    data_files: polymarket/prices/*.csv
    features:
      - name: market_id
        dtype: int64
      - name: slug
        dtype: string
      - name: question
        dtype: string
      - name: category
        dtype: string
      - name: token_id
        dtype: string
      - name: outcome
        dtype: string
      - name: ts
        dtype: int64
      - name: datetime
        dtype: string
      - name: price
        dtype: float64
  - config_name: manifold_prices
    data_files: manifold/prices/*.csv
    features:
      - name: market_id
        dtype: string
      - name: slug
        dtype: string
      - name: question
        dtype: string
      - name: category
        dtype: string
      - name: ts
        dtype: int64
      - name: datetime
        dtype: string
      - name: outcome
        dtype: string
      - name: price
        dtype: float64

StakeBench: Evaluating Language Understanding Grounded in Market Commitment

Version: 1.1.2 (Hugging Face release; the data are those of release 1.1.1 on Google Drive, with two Polymarket files re-encoded to UTF-8, see Hugging Face notes)
Date: 2026-11-04
License: CC BY 4.0
Paper: Pei, Hu, Shi, Ma, Liu, Cartlidge. StakeBench: Evaluating Language Understanding Grounded in Market Commitment. NeurIPS 2026, Track on Evaluations and Datasets. arXiv: https://arxiv.org/abs/2605.26074 · OpenReview: https://openreview.net/forum?id=FWuX7Tx9gK
Code: https://github.com/PEIYUNHUA/stakebench_code_NeurIPS26
Archival copy (release 1.1.1, referenced in the paper): https://drive.google.com/drive/folders/1Xg83tRFH-BghesJUUGWY5XacQliCRYhu

Overview

StakeBench is a large-scale evaluation framework for commitment-grounded language understanding in prediction markets. It links 560,876 comments from 2,261 resolved markets to each author's verified position record, post-comment trading actions, and market price histories across two public platforms — Polymarket (real-money, USD) and Manifold Markets (virtual-currency, Mana).

The key design principle is to replace perception-based supervision with revealed-preference signals derived from observable market commitments: position sides, post-comment trading actions, market resolutions, and price trajectories replace human annotation across four progressive diagnostic tasks.

	Polymarket	Manifold
Comments (cleaned)	514,314	46,562
Unique comment IDs	266,446	46,562
Markets / Topics	2,070	191
Unique users	24,909	3,250
Date range	May 2024 → Mar 2026	Jan 2022 → Mar 2026
Currency	Real USD	Virtual Mana (M$)
Position coverage (G1/G2)	27.2% (139,848)	80.5% (37,430)
Resolved outcome	Price-inferred (≥0.90)	Direct API field

Hugging Face notes

Loading. Each main CSV is exposed as a configuration of this dataset (see the YAML header): polymarket_comments (default), manifold_comments, polymarket_positions, manifold_positions, polymarket_markets, manifold_markets, polymarket_topics, manifold_markets_summary, polymarket_prices, manifold_prices. The raw/ and model_outputs/ folders are plain files; download them with snapshot_download. The two price configurations declare their column types in the YAML header (token_id as a string, because Polymarket token IDs exceed 64-bit integers; category as a string, because it is empty in some files); when reading the CSV files directly, pass dtype={"token_id": str} to pandas.
```
from datasets import load_dataset
comments = load_dataset("greenday123/StakeBench", "polymarket_comments", split="train")

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="greenday123/StakeBench", repo_type="dataset", local_dir="StakeBench")
```
Encoding. On Google Drive, polymarket/comments_clean.csv and polymarket/market_attrs.csv are GB18030-encoded. Here both files are UTF-8 (re-encoded without any other change: same rows, same columns, same values; characters that were outside GB18030 at creation time, mainly emoji, remain ?). The evaluation code in the code repository reads the Drive file with encoding="latin1", so the model inputs in model_outputs/ show non-ASCII punctuation as mojibake (e.g. he¡¯s); the reported results were obtained on that input. For the original text, join raw/polymarket/comments.csv on market_id, comment_id and created_at. All other files are UTF-8 on both hosts.
Checksums. croissant.json lists the SHA-256 of every distributed file for this release; the two re-encoded files differ from the Drive copies, everything else is byte-identical.

Four Diagnostic Tasks

StakeBench is organized as a progressive diagnostic suite. All targets are derived from public records without human annotation.

Task	Input	Prediction	Target	Notes
G1 Position Detection	question, comment	Does commenter hold a position?	`y ∈ {0,1}`	Acc, F1
G2 Belief Inference	question, comment	Which side does commenter's position support?	`z ∈ {YES, NO}` + abstention	DA, CCA, CSD
G3 Language–Action Consistency	question, comment, side, stake	What is user's first post-comment position update?	`a ∈ {flip, increase, decrease, hold}`	Acc₄, Accf, Accd
G4 Belief–Price Lead Detection	window of positioned comments	Does aggregate committed signal predict next price direction?	`d ∈ {UP, DOWN}`	PLA, ΔPLA, n₊

The four tasks form an increasingly demanding hierarchy: detecting financially committed speech → inferring committed belief → predicting subsequent action → detecting collective price-leading signal.

Repository Structure

StakeBench/
├── README.md                         ← this dataset card
├── LICENSE                           ← CC BY 4.0
├── croissant.json                    ← MLCommons Croissant 1.0 metadata (file list with SHA-256, record sets)
├── data_processing.md                ← full data collection & cleaning documentation
│
├── polymarket/
│   ├── comments_clean.csv            ← primary data: 514,314 comment rows (G1–G2)
│   ├── market_attrs.csv              ← market-level metadata (2,070+ markets)
│   ├── topics_summary.csv            ← event/topic-level aggregates
│   ├── positions_clean.csv           ← G3 position-change records (862,539 rows)
│   └── prices/
│       └── {topic_id}.csv            ← hourly price history per topic (G4), 2,055 files
│
├── manifold/
│   ├── comments_clean.csv            ← primary data: 46,562 comment rows (G1–G2)
│   ├── market_attrs.csv              ← market-level metadata (191 markets)
│   ├── markets_summary.csv           ← market-level aggregates
│   ├── positions_clean.csv           ← G3 position-change records (387,262 rows)
│   └── prices/
│       └── {market_id}.csv           ← probability history per market (G4), 191 files
│
├── raw/                              ← unmodified collection outputs (see raw/README.md)
│   ├── README.md
│   ├── polymarket/
│   │   ├── comments.csv              ← comment crawl output, 514,314 rows, UTF-8
│   │   ├── user_positions.csv        ← activity records, 907,375 rows
│   │   └── valid_topics.jsonl        ← 22,240 candidate events with member markets
│   └── manifold/
│       ├── comments.csv              ← comment crawl output, 119,805 rows / 2,926 markets
│       ├── user_positions.csv        ← bet records, 590,904 rows
│       └── valid_markets.jsonl       ← 6,083 candidate resolved binary markets
│
└── model_outputs/                    ← per-sample outputs of the 15 evaluated models (see model_outputs/README.md)

File Descriptions

Primary Benchmark Files (G1 / G2)

`polymarket/comments_clean.csv` (353 MB, 514,314 rows)

Each row is a comment on a Polymarket prediction market with the commenter's reconstructed financial position at comment time.

Field	Type	Description
`market_id`	str	Polymarket market ID
`slug`	str	URL slug
`question`	str	Market question text
`condition_id`	str	On-chain EVM condition ID
`category_validated`	str	Topic category (see §Topic Structure)
`market_start_time`	ISO8601	Market open time (UTC)
`market_end_time`	ISO8601	Market close/settlement time (UTC)
`parent_entity_type_used`	str	API entity type: `Event` or `Series`
`parent_entity_id_used`	str	Entity ID used for comment API call
`comment_id`	str	Comment ID (may repeat for Series-pooled markets)
`created_at`	ISO8601	Comment timestamp (UTC)
`text`	str	Comment body (plain text)
`pseudonym`	str	Platform display name
`user_address`	str	On-chain ETH wallet address (pseudonymous)
`proxy_wallet`	str	Polymarket proxy wallet address
`position_side`	str	`Yes` / `No` / team name at comment time; NaN if no position
`position_size`	float	USD net cost (negative = profitable exit; NaN if no position)
`position_shares`	float	Net shares held at comment time
`reactions`	JSON str	Raw reaction objects array
`reaction_count`	int	Total reaction count (−1 = API anomaly; treat as 0)
`report_count`	int	Number of reports
`parent_comment_id`	str	Parent comment ID (NaN for top-level)
`reply_count`	int	Reply count within collection window

Deduplication note: 48.2% of rows share a comment_id with another row (same comment collected under multiple market_id values for Series-pooled markets). For unique-comment analysis, apply df.drop_duplicates(subset='comment_id', keep='first') to obtain the 266,446 unique comments.

G1/G2 instances: 139,848 rows have position_size > 0 (27.2% of all rows). These form the G1 positive class and the G2 evaluation set.

`manifold/comments_clean.csv` (22 MB, 46,562 rows)

Each row is a comment on a resolved Manifold binary market with position and probability context.

Field	Type	Description
`market_id`	str	Manifold market ID (alphanumeric)
`slug`	str	URL slug
`question`	str	Market question text
`outcome_type`	str	Always `BINARY`
`category`	str	Topic category (see §Topic Structure)
`resolution`	str	Ground-truth resolution: `YES` or `NO`
`comment_id`	str	Unique comment ID
`created_at`	ISO8601	Comment timestamp (UTC)
`text`	str	Plain text (extracted from ProseMirror JSON)
`pseudonym`	str	Username (publicly visible)
`user_address`	str	Manifold user ID
`position_side`	str	`YES` / `NO` at comment time (80.5% coverage)
`position_size`	float	Net Mana M$ invested at comment time
`position_shares`	float	Net shares held at comment time
`prob_at_comment`	float	Market probability snapshot [0,1] at comment time (87.9% coverage)
`reaction_count`	int	Like count
`parent_comment_id`	str	Parent comment ID (NaN for top-level)
`reply_count`	int	Back-calculated reply count

G3 Position-Change Records

`polymarket/positions_clean.csv` (862,539 rows) · `manifold/positions_clean.csv` (387,262 rows)

Timestamped per-trade position records. Used in G3 to derive each commenter's first position update within 48 hours after their comment. Also provides richer position timelines for custom feature engineering.

Polymarket columns: proxy_wallet, timestamp, conditionId, type, side, outcome, outcomeIndex, size, usdcSize, price, transactionHash, title, slug, eventSlug, position_side, position_shares, position_size, yes_cum_shares, no_cum_shares, potential_arbitrage, market_price, position_market_value.

Manifold columns: user_id, username, bet_id, created_at, contract_id, slug, question, outcome, amount, shares, prob_before, prob_after, is_redemption, is_cancelled, is_filled, position_side, position_shares, position_size, yes_cum_shares, no_cum_shares, potential_arbitrage, market_price, position_market_value.

Market Metadata Files

`polymarket/market_attrs.csv`

Full market metadata from the Gamma API. One row per market. Key fields: market_id, question, description, condition_id, clob_token_ids, start_date, end_date, resolved, resolution, volume, liquidity, category_validated, event_id, event_title, event_comment_count.

`manifold/market_attrs.csv`

Full Manifold market metadata. One row per market. Key fields: market_id, question, description, resolution, resolution_probability, probability, volume, unique_bettor_count, comment_count, category_validated.

Summary Files

polymarket/topics_summary.csv: event/topic-level aggregates — event_id, series_id, category, comment_count, total_volume, market_count, event_title, representative_question, best_market_id, best_market_volume, end_date, resolution.

manifold/markets_summary.csv: market-level summary statistics for each of the 191 Manifold markets — market_id, slug, category, outcome_type, resolution, close_time, resolution_time, volume, unique_bettors, total_liquidity, last_comment_time, creator_username, question.

Price History (G4)

polymarket/prices/{topic_id}.csv (2,055 files): hourly CLOB Yes-outcome price per Polymarket topic. Columns: market_id, slug, question, category, token_id, outcome, ts (Unix seconds), datetime (UTC), price (0–1).

manifold/prices/{market_id}.csv (191 files): hourly probability history per Manifold market. Columns: market_id, slug, question, category, ts (Unix seconds), datetime (UTC), outcome, price (YES probability, 0–1).

Topic Structure

StakeBench spans 18 topic–platform combinations: 6 Polymarket topics × 2 platforms, with Manifold split into 12 finer categories.

Polymarket (6 topics): crypto, economics, politics_intl, politics_us, tech_other, other

Manifold (12 topics): crypto, culture, economics, meta, politics_intl, politics_us, science, social, sports_other, tech_ai, tech_other, other

Quick Start

Load comments and filter G1/G2 instances

import pandas as pd

# Polymarket (UTF-8 in this release; the Drive copy needs encoding="gb18030")
poly = pd.read_csv("polymarket/comments_clean.csv", low_memory=False)
poly_dedup = poly.drop_duplicates(subset="comment_id", keep="first")  # 266,446 unique

# G1/G2 instances: position_size > 0
g12_poly = poly_dedup[poly_dedup["position_size"] > 0]   # 139,848 rows (27.2%)
g12_mani = pd.read_csv("manifold/comments_clean.csv")
g12_mani = g12_mani[g12_mani["position_size"] > 0]       # 37,430 rows (80.5%)

G2 — Belief Inference

# Binary YES/NO instances for G2
g2_poly = g12_poly[g12_poly["position_side"].isin(["Yes", "No"])]
g2_mani = g12_mani[g12_mani["position_side"].isin(["YES", "NO"])]

G3 — Language–Action Consistency

# Load position-change records, join to comments by user and market
trades_poly = pd.read_csv("polymarket/positions_clean.csv", low_memory=False)
# For each comment in g2_poly, find first trade within 48h after created_at

G4 — Price Lead Detection

import os, pandas as pd

prices = {}
for fname in os.listdir("polymarket/prices"):
    topic_id = fname.replace(".csv", "")
    prices[topic_id] = pd.read_csv(f"polymarket/prices/{fname}")
# Aggregate positioned comments into time windows; compare to price direction

The task instances evaluated in the paper are built from these files by common.py in the code repository (see Benchmark instance filters).

Dataset Statistics

Core Scale

Metric	Polymarket	Manifold
Comments (cleaned)	514,314	46,562
Unique comments (dedup)	266,446	46,562
Markets	2,070	191
Unique users	24,909	3,250
G1/G2 instances (position_size > 0)	139,848 (27.2%)	37,430 (80.5%)
G3 position-change records	862,539	387,262
G4 price history files	2,055	191

Comment Length (characters)

	p10	p25	p50	p75	p90	max
Polymarket	~10	22	44	85	~160	1,492
Manifold	~18	49	107	230	~470	9,132

Polymarket Stake-Size Tiers (G1/G2 instances)

Tier	Range (USD)	Count	% of positioned
Retail	< $100	52,259	37.4%
Small	$100–$500	35,042	25.1%
Medium	$500–$1,000	14,162	10.1%
Large	$1,000–$5,000	23,865	17.1%
XLarge	$5,000–$10,000	6,315	4.5%
Whale	$10,000–$50,000	6,732	4.8%
Mega	> $50,000	1,473	1.1%

Comments per Market

Metric	Polymarket	Manifold
Median	70	163
Mean	248.5	243.8
Min (filter threshold)	10	100
Max	25,049	990

Manifold Resolution Distribution

Resolution	Comments	Share
YES	25,009	53.7%
NO	21,553	46.3%

Data Collection

All data were collected from public APIs requiring no authentication:

Platform	Data	Endpoint
Polymarket	Markets / tags	`gamma-api.polymarket.com/markets`
Polymarket	Comments	`gamma-api.polymarket.com/comments`
Polymarket	Position activity	`data-api.polymarket.com/activity`
Polymarket	Price history	`clob.polymarket.com/prices-history`
Manifold	Markets	`api.manifold.markets/v0/search-markets`
Manifold	Comments	`api.manifold.markets/v0/comments`
Manifold	Bets (positions)	`api.manifold.markets/v0/bets`

Collection funnel:

Stage	Polymarket	Manifold
Comment-crawl input	2,469	6,083
With collected comments	2,070	2,926
After comment-depth filter	2,070	191

Polymarket keeps markets with ≥ 10 valid comments; Manifold keeps markets with ≥ 100 valid comments (enforcing discussion depth for G3 linkage quality). Collection period: January–March 2026. See data_processing.md for full pipeline documentation.

Raw collection files

raw/ contains the unmodified outputs of the collection scripts for both platforms: the candidate market lists (valid_topics.jsonl, valid_markets.jsonl), the comment crawls (comments.csv) and the per-user position records (user_positions.csv) from which *_clean.csv are derived by collection/clean_comments.py and collection/clean_positions.py in the code repository. raw/README.md lists each file's rows, producer script, columns and the exact cleaning steps.

Benchmark instance filters

comments_clean.csv is the collected corpus after market-level cleaning only (markets with an empty question or fewer than 10 comments removed on Polymarket, fewer than 100 on Manifold). The task instances evaluated in the paper are built from it at run time by common.py in the code repository, with the parameters below (config.py). Comments that fail these filters are retained in the released files so that the tasks can be rebuilt under other thresholds.

Parameter	Value	Where
Minimum comment length after URL stripping (`MIN_TEXT_LEN`)	30 characters	`common.is_quality_text`
Alphabetic-character ratio	≥ 0.30	`common.is_quality_text`
English heuristic (`is_english`)	≥ 85% ASCII letters and ≥ 3% English function words	`common.is_english`
Non-financial market blacklist	wordle / crossword / spelling-bee style puzzle markets	`common._NONFINANCIAL_BLACKLIST`
G3 action window (`HOLD_HOURS`)	48 h after the comment (6–72 h sensitivity in Appendix C)	`common.build_task2_pairs`
G3 hold band (`ACTION_THR`)	position-size change within ±10% of the pre-comment size = `hold`	`common.build_task2_pairs`
G4 near-tie filter (`TASK3_PRICE_EPS`)	windows with \|Δ odds\| ≤ 0.005 removed	`config.py`
G4 window construction (`N_WINDOWS_TARGET`, `MIN_COMMENTS_PER_WINDOW`, `MIN_WINDOWS_PASS`, `TASK3_MIN_WINDOW_COMMENTS`, `TASK3_NEAR50_LOWER/UPPER`)	8 / 10 / 2 / 2 / 0.30–0.70	`config.py`
Consensus baseline (`CONSENSUS_STRENGTH_THR`, `PRICE_EPS`)	consensus direction only if strength ≥ 0.15; \|Δ odds\| ≤ 0.01 counted as no move in the consensus statistics	`config.py`
G1 sampling floor (`MIN_EXPLICIT_TASK0`)	≥ 5 explicit-position comments per class	`config.py`
Instances per topic (`N_MARKETS_PER_TOPIC`, `DEFAULT_TOPIC_CFG`, `DEFAULT_TOPIC_CFG_POLYMARKET`)	per-task counts per topic, e.g. Polymarket default 40 / 80 / 30 / 15 for G1–G4	`config.py`
Decoding	greedy, temperature 0, seed 42	`config.py`

Multilingual headroom

The benchmark is scoped to English. The table reports what the English heuristic removes among quality-passing comments; the removed comments remain in comments_clean.csv. Language identification is heuristic (ASCII-letter ratio and English function-word rate), not a validated language-ID model.

Platform	Quality-passing comments	Kept as English	Removed as non-English	Positioned comments removed
Polymarket	326,162	302,941	23,221 (7.1%)	5,293 (5.7% of 93,398)
Manifold	38,420	37,693	727 (1.9%)	512 (1.7% of 30,826)

The removed Polymarket set is dominated by Romanian, Spanish, Portuguese and French threads on political markets. A non-English track would need the quality filter, the explicit-position lexicon and a verified language-identification step rebuilt for the target language.

Model outputs

model_outputs/ — per-sample predictions of the 15 evaluated models, prompt audit logs, and the Wrong-Bettor / DiD audit inputs. See model_outputs/README.md.
The result files behind Section 5.5 and the appendices added in the camera-ready version (lexical probe, activity-skew stress test, G4 window-level probe and its window pool, multiplicity-corrected Wrong-Bettor audit, verified examples) are not distributed as files; the scripts that regenerate them from model_outputs/ and the benchmark files are in the code repository under analysis/.

Responsible AI Statement

Data Limitations

Polymarket G1/G2 position coverage is 27.2%: Many users comment without holding a position. Tasks requiring position labels should account for this imbalance. Use position_size > 0 as the inclusion criterion for G1/G2 (not merely position_side non-null).
Polymarket comment deduplication: 48.2% of rows are duplicates across Series-pooled markets. Analyses must deduplicate on comment_id before unique-comment tasks.
Temporal non-stationarity: Both platforms show strong temporal spikes (e.g., Polymarket peaks in June 2025 due to Iran/nuclear markets). Models trained without temporal splits may overfit to event clusters.
Manifold uses virtual currency: position_size values are in Mana (M$), not comparable to Polymarket USD. Cross-platform financial comparisons require normalization.
English language bias: The vast majority of comments are in English. Non-English markets exist but are underrepresented; the English heuristic removes 7.1% of quality-passing Polymarket comments and 1.9% on Manifold (see Multilingual headroom).
G3 label coverage on Polymarket: G3 labels exist for 73,893 of the 139,848 positioned Polymarket comments, and for 27,116 of 37,430 on Manifold. A comment whose author has no reconstructed position record cannot receive a G3 label, and a further share of positioned comments is removed by the text-quality filter (see Benchmark instance filters). Whether the unreconstructed Polymarket records differ systematically from the retained ones cannot be tested from the released data.

Data Biases

Topic concentration: Polymarket is heavily weighted toward US/global politics. Sports and economics are underrepresented.
User concentration: Top users account for a disproportionate share of comments (power-law distribution). Analyses may be dominated by prolific users.
Platform self-selection: Prediction market users skew toward financially engaged, English-speaking users with interest in forecasting — not a representative sample of the general population.
Commitment-non-commitment selection: Position-holding commenters differ from non-holders on observable dimensions with effect sizes below d = 0.3 (Cohen's d), indicating modest but non-zero selection bias in G2 and G3 evaluation sets.

Sensitive Information

On-chain wallet addresses: Polymarket user_address and proxy_wallet are public Ethereum addresses. They are pseudonymous but potentially re-identifiable via on-chain analysis. Researchers should not attempt to de-anonymize users.
Display names: Polymarket pseudonym and Manifold pseudonym are platform-assigned or user-chosen display names, publicly visible on the respective platforms.
No real names, IP addresses, or private messages are included.
The raw files in raw/ contain the same identifier types and no others; see raw/README.md.

Construct Validity

position_side as belief proxy: Revealed position side is a behavioral proxy for private belief, not a direct observation. It reflects observable market commitment, not an unobserved mental state.
G3 action labels: Derived from net position changes in the trading record within 48 hours. Labels are deterministic reconstructions from public records, not self-reported intentions.
Market resolution as ground truth: Manifold resolution (YES/NO) is human-adjudicated per platform resolution policy. It reflects platform rules, not necessarily objective ground truth.

Potential Misuse

Deanonymization: Combining user_address with public blockchain data could reveal user identity. This is discouraged and may violate platform terms of service.
Automated content generation: The dataset should not be used to train models that generate fake prediction market comments for manipulation purposes.

Synthetic Data

This dataset contains no synthetic data. All comments, positions, and prices are sourced directly from Polymarket and Manifold public APIs.

Provenance

Data collected under Polymarket's Terms of Use and Manifold's Terms of Service, both of which permit non-commercial research use. Collection and cleaning scripts are in the code repository under collection/; their unmodified outputs are in raw/.

Benchmark Results (15 LLMs, 18 topic–platform splits)

Key findings from the StakeBench evaluation (camera-ready version of the paper):

G1 Position Detection: 11 of 15 models exceed 0.65 accuracy (best Qwen3-32B, 0.747); model rankings transfer across platforms (ρ = 0.975)
G2 Directed Accuracy: 0.506–0.599 across 15 models (closed-source avg. 0.593, finance-domain avg. 0.527)
G3 class collapse: 10 of 15 models collapse to one or two action labels; mean hold recall 22.9% although hold is 47.6% of true actions
G4: all ΔPLA values are negative (−0.022 to −0.437); no model consistently exceeds the stake-weighted majority baseline
Scale: model parameter count is at most weakly associated with any task score across the 3B–32B open-weight range (|ρ| ≤ 0.43)
Finance-domain tuning: associated with weaker speaker-level commitment comprehension than comparable general models (mean G3 action accuracy 0.171 vs. 0.271 for closed-source models)

See the paper for full tables and analysis.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to share and adapt this data for any purpose, including commercial use, provided you give appropriate credit, provide a link to the license, and indicate if changes were made.

See LICENSE for the full license text.

Citation

If you use StakeBench in your research, please cite:

@inproceedings{pei2026stakebench,
  title     = {StakeBench: Evaluating Language Understanding Grounded in Market Commitment},
  author    = {Pei, Yunhua and Hu, Jingyu and Shi, Yiwei and Ma, Hongnan and Liu, Weiru and Cartlidge, John},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
  year      = {2026},
  note      = {arXiv:2605.26074}
}

Changelog

1.1.2 (2026-11-04, Hugging Face release): polymarket/comments_clean.csv and polymarket/market_attrs.csv re-encoded from GB18030 to UTF-8 (no other change); croissant.json regenerated with the new checksums; dataset card: corrected the column lists of the price files and of positions_clean.csv, updated the Benchmark Results summary to the camera-ready numbers, added the Hugging Face loading notes. All other files are byte-identical to 1.1.1.
1.1.1 (2026-10-30, Google Drive): croissant.json rebuilt to Croissant 1.0 with per-file checksums and record sets.
1.1 (2026-10-27): de-anonymised metadata; corrected the G3 coverage statement; added Benchmark instance filters, Multilingual headroom, raw/ (unmodified collection outputs) and model_outputs/; added the encoding note for polymarket/comments_clean.csv; corrected the positions_clean.csv row counts (header line had been counted). Data files are byte-identical to 1.0.
1.0 (2026-05): submission version.