▌ Open data · CC-BY 4.0
Raw votes.
Every vote cast on the PlotPoints benchmark, served as CSV. Re-run the analysis, build your own ranking, or audit ours. Each vote carries its voter's random id, so the vote-stuffing checks can be re-run too. IP hashes and user agents are withheld; see What's excluded below for the full list.
▌ Downloads
Issue № 04 · September 2026
Does it know where the line is?
This round's votes are held back while its arena is open. The file fills once voting closes.
Issue № 02 · April–June 2026
Single-turn charm doesn't survive twelve turns.
1,943 votes · 482 voters
Issue № 01 · April 2026
The champ that wouldn't move.
1,857 votes · 335 voters
Everything
All rounds, flattened.
Every vote across every round in one CSV. Filter by the
round column.▌ Column schema
| Column | Meaning |
|---|---|
| id | UUID per vote |
| round | 1, 2, 3, 4 · which round bucket |
| mode | arena · multiturn_arena · nsfw_multiturn_arena · ladder_arena · rubric |
| scenario_id | Pair id (arena modes) or response id (rubric). Votes from a blind arena (ladder_arena) are left out of the export while its round is open and published in full once it closes |
| context | Underlying seed/scene id |
| model_a, model_b | The two models in an arena matchup, in the order shown (A left, B right) |
| winner | A · B · tie (arena modes only) |
| model | Model that produced the response (rubric only) |
| scores | JSON. Rubric: { axis_id: 1-5 }. Multi-turn arenas (Round 04 on): { dwell_ms, export_id, export_id_mismatch } when known; export_id_mismatch is true when the voter's page was from another export |
| notes | Voter free-text (rubric, optional) |
| is_catch | true if the matchup was a catch-pair calibration |
| catch_correct | true if the voter picked the pre-declared winner |
| source | 'native' for plotlight, 'arena_l3vi4th4n_round_N' for imports |
| signed_in | 1 if cast by an authed user; 0 if anonymous |
| client_timestamp | Wall-clock at submit (best-effort) |
| created_at | Server insert time (authoritative) |
| voter_id | Random UUID per voter, set once per browser; it carries no account, IP or device data. Every vote from one browser shares it, so grouping by it reproduces the vote-stuffing and voter-quality checks |
▌ What's excluded
- user_idAuth binding · surfaced as the `signed_in` boolean only.
- ip_hashSpam/abuse signal · withheld for privacy.
- user_agentPII-adjacent · withheld for privacy.
▌ Upstream / canonical
The CSV here is what Plotlight collected. The full benchmark dataset, including model outputs, judge rationales, and the analyzer scripts that turn raw votes into the standings, lives on HuggingFace as the canonical artifact: