Datasets:
Leaderboard question: the current #1 was trained on test-split rows (Phocinae-Largha-150M-v1)
Hi maintainers, and congratulations to the Phocinae team. A 144M model at 21 ms per decision is genuinely impressive.
While reading the Phocinae-Largha-150M-v1 model card, we found this in its evaluation-overlap disclosure (corrected 2026-10-10): the fine-tune mix includes a 500-item error-correction pool, with gold soft targets, "mined from model predictions on the typed-decisions test rows".
Credit to the authors for disclosing it openly. Still, it means the 0.906 at the top of the Hub leaderboard was measured on a test split the model has partly studied, answer distributions included. Even a spotted seal swims faster in water it has already mapped. 🦭
Three short questions, asked in good faith:
- Should entries with test-split exposure be flagged on the Hub leaderboard and in Table 2, or listed separately? Table 2 already separates "fitted on train" from zero-shot, and "fitted on test" seems like a third category.
- Would it help to add one line to the dataset card, for example: "Do not train on test rows, test labels, or pairs derived from test predictions", and to ask submitters to state this in their
.eval_resultsnotes? - @perrylink , would you consider re-scoring a checkpoint trained without the test-derived pool? That number would be interesting in its own right, and it would make the comparison fair for everyone.
Thanks for running this benchmark. Clean splits are what make the numbers worth chasing.
Hi — thanks for reading the card that closely, and for asking in the open. Short answers, in the order you asked, plus a couple of commitments.
The disclosure is accurate, and it went in on purpose. We found the overlap while auditing our own pipeline and corrected the model card on 2026-10-10 — including an earlier line of ours that was simply wrong. You're right about what it implies: the 0.906 was measured on a split this lineage has partly seen, answer distributions included, and we don't treat it as a zero-shot result. Our entry carries the "fitted on the train split" framing; the pair pool is the part added on 10-10 — now impossible to miss on the card.
Flagging — count us in favour of a third category, and please mark our entry with it as soon as a mechanism exists ("fitted on train, test-exposed", or whatever reads cleanest to you). Between a flattering table and an accurate one, we'll take accurate; that's what the disclosure was for. The mechanism is the maintainers' call — we'll follow their lead.
Dataset card + notes — agreed on both. We've just updated our
.eval_resultsnotes with the overlap statement, so the board entry carries the caveat at the source. A line in the dataset card like "do not train on test rows, test labels, or pairs derived from test predictions" sounds right to us, and asking submitters to state this in their notes is a reasonable default. If a ready-to-paste wording would help, we're happy to draft one from ours.A clean re-score — yes, that's what we want from this lineage too, and it's already in motion: same recipe minus the test-derived pairs, same 400 cases, and we'll post the number here next to the current one — high or low. Aiming for weeks, not months; we'd rather be slow and honest than fast and wrong. Full overlap details stay in the model card's disclosure section either way.
Thanks for running this benchmark, and for keeping the bar honest — clean splits are the only reason any of these numbers mean anything. (And a fair line about the seal. 🦭)
— Perry (Phocinae)
Hi all — promised follow-up: the clean re-score is done. Numbers first, then the full picture.
Setup. Same recipe as the released v1.1 model, minus the 500 test-derived error-correction pairs: they were swapped back out in favour of standard base items from the same source and seed (dose kept constant), with everything else identical — same two-stage fine-tune, same starting checkpoint, epochs, seeds, and evaluation. Same 400 cases / 2,000 decisions. Two reads of the final checkpoint matched exactly (0.791 = 0.791).
Result. Clean re-score, typed-decisions: 0.791 vs 0.906 as submitted.
Where the difference sits. All of it is on the mined rows. On the 500 decisions whose test rows fed the pair pool: 0.198 clean vs 0.724 as submitted. On the other 1,500: 0.989 clean vs 0.967 as submitted. So the pool's benefit was almost entirely a test-row effect; off those rows the model is unchanged-to-slightly-better.
The rest of the battery, same honesty pass. Option-order flip (reversed), 150/400: 0.0367 / 0.0250 clean vs 0.0200 / 0.0217 as submitted. JevBench public-231: 0.5195 clean vs 0.5455 as submitted. Everything, high or low, as measured.
What's now in place. Both figures live on the card, the board entry's .eval_results notes, and BENCHMARKS.md. The replacement items were re-scanned against the evaluation sets before training, with no 60-character substring matches, and the re-score was produced and read fresh end to end. The 0.906 entry stands as measured and disclosed; as offered, we'd still support flagging it (or a companion-value mechanism) if the maintainers want one. Happy to write the re-score setup up as a short reproducible note too — this class of check should be cheap for anyone's fine-tune.
Thanks again for raising it in good faith. It made our pipeline better.
— Perry (Phocinae)
Thanks @perrylink for the clean re-score and the candor — 0.791 with the test-derived pairs removed, and the per-row breakdown showing the gain sat almost entirely on the mined rows. That is the honest number the board needs.
Maintainers: now that a clean, same-recipe figure exists (0.791), could the leaderboard and Table 2 reflect it for ranking — either rank on the clean / zero-shot-comparable value, or add the “fitted on train, test-exposed” flag so the 0.906 entry reads correctly next to clean entries? The disclosed 0.906 can remain as a measured companion value; the ranking is where the test exposure actually matters.
Happy to help with the dataset-card line (“do not train on test rows, test labels, or pairs derived from test predictions”) if useful. Clean splits are the whole point. 🦭