Datasets:

hayshn
/

ICAIF26-EventXbench

@@ -1,230 +1,105 @@
-# EventX: A Multimodal Benchmark Linking Social Media Posts to Prediction Market Dynamics
-**Dataset: [https://huggingface.co/datasets/mlsys-io/EventXBench](https://huggingface.co/datasets/mlsys-io/EventXBench)**
-[![Dataset on HF](https://img.shields.io/badge/%F0%9F%A4%97%20Dataset-EventXBench-blue)](https://huggingface.co/datasets/mlsys-io/EventXBench)
-[![Paper](https://img.shields.io/badge/Paper-ACM%20MM%20'26-red)](https://doi.org/PLACEHOLDER)
-[![License: CC BY-NC 4.0](https://img.shields.io/badge/License-CC%20BY--NC%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc/4.0/)
-**EventX** is a multimodal benchmark connecting 9M Twitter/X posts from 1,152 KOL accounts to 11,952 Polymarket prediction markets (2021--2026). It defines seven tasks across two tiers: **resolution** (human-annotated ground truth) and **forecast** (deterministic labels from post-publication tick data).
 ## Quick Start
 ```bash
-pip install -r requirements.txt
-from eventxbench import load_task
-train, test = load_task("t1")  # Returns pandas DataFrames
-# Run a baseline
-python baselines/t1/llm_baseline.py --provider openai --model gpt-4o --shots 0
-# Evaluate predictions
-python evaluation/evaluate.py --task t1 --predictions results/t1_predictions.jsonl
-```
-## Benchmark Overview
-| Task | Name | Tier | Output | Instances | Primary Metrics |
-|------|------|------|--------|-----------|-----------------|
-| T1 | Market Volume Prediction | Forecast | 3-class (`high`/`moderate`/`low`) | 305 | Macro-F1, `high`-class P@K |
-| T2 | Post-to-Market Linking | Resolution | Market ID or `none` | 815 | Accuracy@1, MRR |
-| T3 | Evidence Grading | Resolution | Ordinal 0--5 | 342,552 | QWK (kappa), macro-F1 |
-| T4 | Market Movement Prediction | Forecast | Direction x Magnitude | 4,803 | Dir-Acc, Mag-F1, Spearman rho |
-| T5 | Volume & Price Impact | Forecast | Continuous | 407 (268 clean) | Spearman rho (price_impact, volume_multiplier) |
-| T6 | Cross-Market Propagation | Forecast | 3-class | 4,006 | Macro-F1, MAE (onset lag) |
-| T7 | Impact Persistence (Decay) | Forecast | 3-class (`transient`/`sustained`/`reversal`) | 407 (268 clean) | Macro-F1 |
-## Tasks
-### T1: Conditional Market Volume Prediction
-Predict the final trading-volume percentile of a subsequently created market, given pre-market social signals.
-- **Labels**: `high` (>80th pctl), `moderate` (40th--80th), `low` (<40th)
-- **Input**: Tweet cluster features, event metadata, temporal features
-- **Metrics**: Macro-F1, `high`-class precision@K
-### T2: Post-to-Market Linking
-Given a tweet and a candidate set recalled by BGE-large-en-v1.5 / FAISS dense retrieval, identify which market the post addresses (or `none`).
-- **Input**: Tweet text + candidate market questions
-- **Metrics**: Accuracy@1, MRR, `none`-class F1
-### T3: Evidence Grading and Resolution Potential
-Assign an ordinal evidence grade (0--5) to each post-market pair.
-- **Grade scale**: 0 (`noise`), 1 (`commentary_reaction`), 2 (`speculation_rumor`), 3 (`indirect_report`), 4 (`strong_direct`), 5 (`resolving`)
-- **Metrics**: Quadratic-weighted kappa, `resolving`-class precision, macro-F1
-### T4: Market Movement Prediction
-Predict direction and magnitude of the YES-price change at 2-hour horizon after a tweet.
-- **Direction**: `up` (delta > 0.02), `down` (delta < -0.02), `flat` (otherwise)
-- **Magnitude**: `large` (>8%), `medium` (2--8%), `small` (<=2%)
-- **Metrics**: Direction accuracy, Magnitude macro-F1, Spearman rho on continuous delta curve
-- **Secondary horizons**: 30 min, 6 h
-### T5: Volume and Price Impact
-Predict two continuous targets per post: (i) `price_impact` (max absolute deviation from p0), (ii) `volume_multiplier` (total volume / 24h baseline).
-- **Targets**: `price_impact` (continuous), `volume_multiplier` (continuous)
-- **Metrics**: Spearman rho for each target
-- **Note**: T5 and T7 share the same underlying data (`task5+7/` in the codebase). T5 evaluates the continuous predictions; T7 evaluates the decay classification.
-### T7: Impact Persistence (Decay)
-Classify whether a tweet's initial market impact is transient, sustained, or reverses over time.
-- **Labels**: `transient` (|delta_2h| < 30% of |delta_15m|), `sustained` (same sign, larger), `reversal` (sign flips at 2h)
-- **Metrics**: Macro-F1
-- **Note**: Uses the same data as T5 but evaluates the `decay_class` field. In the codebase, uses `task5+7/` directories and `t7_` prefixes.
-### T6: Cross-Market Propagation
-Predict whether a tweet's market impact propagates to sibling markets within 2 hours. A sibling is deemed "moved" if |delta_p| > 1.5 sigma (rolling 24h stdev).
-- **Labels**: `no_effect`, `primary_mover` (primary market moves first, sibling follows), `propagated_signal` (sibling moves before or instead of primary)
-- **Metrics**: Macro-F1, MAE on onset lag (minutes)
-## Data
-### Hugging Face Dataset
-All data is hosted on Hugging Face:
-```python
 from datasets import load_dataset
-ds = load_dataset("mlsys-io/EventXBench", "t1")
-# Or use our convenience loader
-from eventxbench import load_task
-train, test = load_task("t1")
 ```
-See [`data/README.md`](data/README.md) for the full dataset card.
-### Data Files
-| File | Description | Size |
-|------|-------------|------|
-| `posts_no_text.jsonl` | Tweet IDs and metadata (text stripped for privacy) | ~9M rows |
-| `market_fundamental.json` | Market metadata (question, category, resolution) | 11,952 markets |
-| `market_ohlcv.json` | Price/volume time series (OHLCV) | 1.8 GB |
-| `t1_labels.jsonl` | T1 ground truth with train/test splits | 305 |
-| `t2_groundtruth.jsonl` | T2 post-market linking pairs | 815 |
-| `t3_graded.json` | T3 evidence grades (0--5) | 342,552 |
-| `t4_labels.jsonl` | T4 direction x magnitude labels | 4,803 |
-| `t5_labels.jsonl` | T5 price impact + volume multiplier | 407 |
-| `t6_labels.jsonl` | T6 cross-market propagation labels | 4,006 |
-| `t7_labels.jsonl` | T7 decay class labels (same data as T5) | 407 |
-### Privacy
-Tweet text is **not** included in the public release to comply with Twitter/X Terms of Service. The `posts_no_text.jsonl` file contains tweet IDs for rehydration via the Twitter API. Market data from Polymarket is fully included.
-## Baselines
-We provide three baseline families:
-### LLM Baselines
-Zero-shot and few-shot prompting:
-- GPT-4o (OpenAI)
-- Sonnet 4.5 (Anthropic)
-- Grok 4.1 (xAI)
-- Qwen 3.5 (local via vLLM, 4B and 27B)
-```bash
-# T1 zero-shot with GPT-4o
-python baselines/t1/llm_baseline.py --provider openai --model gpt-4o --shots 0
-# T4 three-shot with Claude
-python baselines/t4/llm_baseline.py --provider anthropic --model claude-sonnet-4-5-20250514 --shots 3
-# T2 contextual validation with an OpenAI-compatible Lumid model
-python baselines/t2/llm_baseline.py \
-  --provider openai \
-  --base-url https://lum.id/llm \
-  --api-key-env LUMID_API_KEY \
-  --model nvidia/Gemma-4-26B-A4B-NVFP4 \
-  --disable-thinking \
-  --split val \
-  --shots 0 \
-  --resume \
-  --output results/t2_gemma.val.0shot.jsonl
-```
-**Required env vars**: `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `XAI_API_KEY`, `HF_TOKEN` (for Qwen)
-The T2 runner defaults to the frozen 2,500-row contextual-gold package, treats
-`NONE` as an explicit option, and writes Accuracy@1, MRR, NONE-class F1,
-bootstrap intervals, token usage, latency, input hashes, and prompt provenance.
-Run validation first. A frozen test run requires `--split test --allow-test`.
-Models recorded as Task 2 train-silver judges are automatically marked
-`diagnostic_excluded_silver_judge` and must not be reported as independent
-paper baselines.
-Lumid also accepts the Anthropic message format. Use `--provider anthropic`
-with the same gateway root and PAT environment variable. The provider flag
-describes the request format; the actual baseline model remains the model ID
-passed with `--model` and returned by Lumid's `/llm/v1/models` endpoint.
-### ML Baselines
-LightGBM classifiers with Bayesian hyperparameter tuning (Optuna):
-```bash
-python baselines/t1/lightgbm_baseline.py
-python baselines/t4/lightgbm_baseline.py
-```
-### Heuristic & Basic Baselines
-Majority class, random walk, graph heuristics, BM25 retrieval:
-```bash
-python baselines/t1/basic_baseline.py
-python baselines/t6/basic_baseline.py
 ```
-## Evaluation
-```bash
-python evaluation/evaluate.py --task t1 --predictions results/t1_preds.jsonl
-python evaluation/evaluate.py --task t4 --predictions results/t4_preds.jsonl
-python evaluation/evaluate.py --task all --predictions-dir results/
-```
-See [`evaluation/README.md`](evaluation/README.md) for prediction format specs.
-## Repository Structure
-```
-EventXBench/
-├── README.md
-├── LEADERBOARD.md
-├── requirements.txt
-├── eventxbench/                  # Data loading utilities
-│   ├── __init__.py
-│   └── loader.py
-├── data/
-│   └── README.md                 # Hugging Face dataset card
-├── baselines/
-│   ├── t1/ ... t6/               # Per-task baselines (LLM, ML, basic)
-│   └── t7/                       # Impact Persistence (Decay) baselines
-├── evaluation/
-│   ├── evaluate.py               # Unified evaluation CLI
-│   ├── metrics.py                # Metric implementations
-│   └── README.md
-├── scripts/
-│   └── upload_to_hf.py           # Upload data to Hugging Face
-└── examples/
-    └── quickstart.py
 ```
-## Leaderboard
-See [LEADERBOARD.md](LEADERBOARD.md) for current results. To submit, open a pull request.
 ## License
-- **Code**: MIT License
 - **Data**: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)
 - Tweet text excluded; use Twitter API for rehydration
 - Polymarket data included under fair use for research

+---
+license: cc-by-nc-4.0
+language:
+- en
+pretty_name: ICAIF26-EventXbench
+size_categories:
+- 1M<n<10M
+task_categories:
+- text-classification
+- tabular-classification
+task_ids:
+- multi-class-classification
+tags:
+- prediction-markets
+- social-media
+- multimodal
+- financial-nlp
+- twitter
+- polymarket
+---
+# ICAIF26-EventXbench
+**EventX** is a multimodal benchmark connecting ~9M Twitter/X posts from 1,152 KOL accounts to 11,952 Polymarket prediction markets (2021–2026). It defines seven tasks across two tiers: **resolution** (human-annotated ground truth) and **forecast** (deterministic labels from post-publication tick data).
+This repository hosts the canonical **KDD v2** release of the EventXBench data (tasks T1, T2, T4, T5, T6). The companion codebase is available at [github.com/mlsys-io/EventXBench](https://github.com/mlsys-io/EventXBench); the primary dataset mirror is [mlsys-io/EventXBench](https://huggingface.co/datasets/mlsys-io/EventXBench).
 ## Quick Start
 ```bash
+pip install datasets
 from datasets import load_dataset
+# T1: train + test (2-way temporal split)
+ds = load_dataset("hayshn/ICAIF26-EventXbench", "t1")
+# T4/T5/T6: train + validation + test
+ds = load_dataset("hayshn/ICAIF26-EventXbench", "t4")
 ```
+Alternatively, use the repository loader (`eventxbench.loader`) with `local_dir=` pointing at this dataset's data directory.
+## Benchmark Overview (v2 counts)
+| Task | Name | Tier | Output | Instances (v2) | Primary Metrics |
+|------|------|------|--------|----------------|-----------------|
+| T1 | Market Volume Prediction | Forecast | 3-class (`high_interest`/`moderate_interest`/`low_interest`) | 984 | Macro-F1 |
+| T2 | Post-to-Market Linking | Resolution | Market ID or `NONE` | 5,543 (5,000 human-gold eval) | Accuracy@1, MRR |
+| T4 | Market Movement Prediction | Forecast | Direction × Magnitude | 10,934 | Dir-Acc, Mag-F1 |
+| T5 | Volume & Price Impact | Forecast | Continuous + `decay_class` | 3,342 | Spearman ρ, Decay macro-F1 |
+| T6 | Cross-Market Propagation | Forecast | 3-class (`no_effect`/`primary_only`/`cross_market`) | 4,583 | Macro-F1 |
+> **T3 (Evidence Grading)** is not included in this repository. See the primary mirror for the rebuilt T3 artifacts (train + gold).
+## Splits
+| Task | Train | Validation | Test | Split policy |
+|------|-------|------------|------|--------------|
+| T1 | 709 | — | 275 | Event- and market-group atomic, temporal 60/0/40 |
+| T2 | 543 | 2,500 | 2,500 | Train: silver + single-human; val/test: three-reviewer human gold, zero cross-split overlap |
+| T4 | 2,875 | 2,268 | 5,791 | Event-cluster atomic temporal (2025-05→2026-06) |
+| T5 | 889 | 692 | 1,761 | Same temporal policy as T4 (non-flat subset) |
+| T6 | 766 | 1,225 | 2,592 | Event-cluster atomic temporal × horizons {1,3,7}d |
+## Task Descriptions
+- **T1 – Market Volume Prediction**: predict the creation-cohort volume percentile of a market from pre-market social signals (post links within 7 days, cosine+entity-verified).
+- **T2 – Post-to-Market Linking**: given a post and a frozen BGE-M3 candidate set (top-10), select the addressed market or `NONE`.
+- **T4 – Market Movement Prediction**: predict the YES-price direction (`up`/`down`/`flat`) and magnitude bucket at 1/3/7-day horizons for a market-day bundle.
+- **T5 – Volume & Price Impact**: predict continuous `price_impact` and `volume_multiplier` targets and the `decay_class` (`transient`/`sustained`/`reversal`).
+- **T6 – Cross-Market Propagation**: predict whether a bundle's move propagates to sibling markets (graph v2, 35,526 edges / 14,739 nodes), labeled `no_effect`/`primary_only`/`cross_market` at 1/3/7-day horizons.
+## Label Distributions
+- **T1**: `high_interest` 480, `moderate_interest` 387, `low_interest` 117
+- **T2**: train LINK 402 / NONE 141; validation LINK 1,179 / NONE 1,321; test LINK 1,198 / NONE 1,302
+- **T4**: `flat` 7,592; `up` 1,730; `down` 1,612 (direction)
+- **T5**: `sustained` 1,932; `reversal` 807; `transient` 300; (303 rows have no computable decay class)
+- **T6**: `cross_market` 2,796; `no_effect` 1,366; `primary_only` 421
+## Leakage Control
+All forecast tasks (T1, T4, T5, T6) use **event-disjoint temporal splits with an embargo ≥ the maximum label horizon**. Features are computed as-of the decision time; labels are strictly in the future. T4/T5/T6 attach `confound_flag` when a different-author high-evidence post appears within the label window. T2's val/test sets are sealed human gold with zero instance, tweet, or exact-text overlap across splits.
+## Privacy
+Raw tweet text is **not** included for non-T2 tasks; records carry `tweet_ids` for rehydration. T2's val/test include post text under the dataset's Data Usage Agreement (ID-only distribution otherwise). Market data from Polymarket is fully included.
+## Files
 ```
+data/
+├── t1/    train.jsonl, test.jsonl
+├── t2/    t2_train.jsonl, t2_val.jsonl, t2_test.jsonl
+├── t4/    train.jsonl, validation.jsonl, test.jsonl
+├── t5/    train.jsonl, validation.jsonl, test.jsonl
+└── t6/    train.jsonl, validation.jsonl, test.jsonl
 ```
 ## License
 - **Data**: [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)
 - Tweet text excluded; use Twitter API for rehydration
 - Polymarket data included under fair use for research