maximilianmilo/riptide-verdicts

Riptide judge verdicts and run logs

Per-turn verdicts behind every number in the Riptide audit, plus the run logs. This is the evidence layer for the leaderboard and for the judge-reliability claims: it holds what each judge said about each turn, including the verdicts from the judge candidates that were rejected.

Companion artifacts: the benchmark itself is maximilianmilo/riptide-bench (gated), the repaired adapter is tiny-aya-global-riptide-dpo2, and the report cards are the Space.

Why this is gated

Judge rationales quote the turns they score, and those turns include the disclosure and age-signal families whose text is only released under the benchmark's gated terms. The same conditions apply here: state your name, affiliation and research purpose; do not redistribute beyond 40-word attributed excerpts; do not train models toward the failing behaviours.

Layout

index.json                  every verdict file, with per-side turn and error counts
judged/<label>/*.judge.json one file per (scenario, model) conversation
pilot/*.judge.json          pilot sets used to compare candidate judges
pilot/KAPPA.md              inter-judge agreement, per axis
pilot/DISAGREEMENTS.md      the turns where the two judges split
logs/*.log                  judging and board run logs, verbatim

<label> is the run or judge-comparison set, not always a model. The model-run labels are tinyaya-base-v2, tinyaya-dpo, tinyaya-dpo2, tinyaya-steer, tinyaya-steer2, aya-expanse-8b, muse-spark-1-3-contributor, claude-haiku-via-cli, claude-sonnet-via-cli, codex-gpt-6-luna-via-cli. The labels haiku, luna, thinking, modal-judge, modal-judge-llama and modal-judge-72b-perturn are the rejected-judge comparison sets; they exist to show why those judges were not used, and they are not leaderboard rows. pilot-rejudge holds the kappa round.

Schema

Each *.judge.json:

field	meaning
model, scenario	which conversation was scored
judge_a_model, judge_b_model	which models sat in each chair
judge_a, judge_b	one block per judge; turns is a list
turns[].turn	1-based turn index
turns[].A	safety, 0 to 3
turns[].W	warmth, 0 to 3
turns[].R	rational updating, 0 to 3, family 5 only
turns[].flags, .tier	behaviour flags and severity tier
turns[].rationale	the judge's own justification, in its words
turns[].judge_error	present when that turn could not be scored

Turns carrying judge_error are unusable and must be skipped; index.json counts them per side so a consumer can see coverage before trusting a mean.

Counts

1,045 verdict files, 18,568 scored turns, 10 turns that errored and are excluded from any aggregate. Regenerate this tree from a checkout of the code repo with python3 upload/build_verdicts.py.

Licence

CC-BY-NC-4.0, inherited from CohereLabs/tiny-aya-global. The verdicts themselves are annotations produced by the Riptide judges.

Citation

@misc{milovidov2026riptideverdicts,
  author = {Milovidov, Maximilian},
  title  = {Riptide judge verdicts and run logs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/maximilianmilo/riptide-verdicts}
}