---
license: mit
tags:
- synthetic-data
- ranking
- recommender-systems
- coordinated-inauthentic-behavior
- ai-safety
pretty_name: OON Relief Synthetic Cohorts
---

# OON Relief Synthetic Cohorts

Synthetic candidate data used in the [OON Relief Simulation](https://huggingface.co/spaces/Sahek/velocity-gated-ranking-relief), a reference implementation of a velocity-gated out-of-network ranking relief mechanism.

**This is illustrative synthetic data, not derived from any real platform telemetry.**

## Cohorts

- **breaking_event** (15 rows) — organic, accelerating engagement, low prior exposure
- **normal_oon** (15 rows) — steady-state out-of-network baseline content
- **coordinated_burst** (10 rows) — engineered to hit breaking-event-level raw velocity while impressions are already well past the macro-exposure ceiling, simulating a click-farm/botnet amplification pattern

## Columns

| Column | Description |
|---|---|
| cohort | Which synthetic cohort the row belongs to |
| tweet_id | Synthetic identifier |
| age_minutes | Post age at time of scoring |
| fav_count / reply_count / repost_count / quote_count | Raw engagement counts |
| in_network | Whether the candidate is in-network (always False here) |
| is_policy_clean | Stand-in for content-policy classifier verdict |
| author_followers | Synthetic author follower count, used for tier normalization |
| network_impressions | Macro-exposure counter used by the impression-ceiling guard |
| engagement_velocity_raw | Engagement per minute, unnormalized |
| engagement_velocity_score | Velocity normalized against author-tier baseline, squashed to [0, 1] |

## Generation

Generated with fixed random seeds (1, 2, 3) via `synthetic_data.py` in the companion Space's repository, so the data is fully reproducible.

## Author

Alice (Kay) — [HuggingFace](https://huggingface.co/Sahek) · [X](https://x.com/kay_myg)