Table 2: list Phocinae-Largha-150M-v1 at its clean re-score (0.791), drop the test-exposed 0.906 (discussion #25)

README.md

@@ -101,11 +101,11 @@ model is learning the teacher's quirks.
 
 | Model | Kind | Accuracy ↑ | KL from gold ↓ | Brier ↓ | ECE ↓ | Reported latency |
 |---|---|---|---|---|---|---|
-| [Phocinae-Largha-150M-v1](https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1) (mmBERT-small, 144.3M) | specialist | 0.906 | 0.0548 | 0.0248 | 0.2519¶ | 21.0 ms‡ (RTX 5090) |
 | [OpenDecider-large-td](https://huggingface.co/manjunathshiva/opendecider-large-td) (Qwen3-Next-80B-A3B + LoRA) | general, then fine-tuned on `train` | 0.801 | 0.081 | 0.044 | – | 440 ms‡ |
 | [od1-typed-decisions](https://huggingface.co/mvbalaji/od1-typed-decisions) (Qwen3.5-4B) | specialist | 0.797 | 0.082 | 0.045 | 0.156¶ | – |
 | [OpenDecider-nano](https://huggingface.co/manjunathshiva/opendecider-nano) (400M encoder) | general, then fine-tuned on `train` | 0.796 | 0.079 | 0.043 | – | 17 ms‡ |
 | [OpenDecider-small-td](https://huggingface.co/manjunathshiva/opendecider-small-td) (Qwen3-4B + LoRA) | general, then fine-tuned on `train` | 0.792 | 0.080 | 0.043 | – | 40 ms‡ |
+| [Phocinae-Largha-150M-v1](https://huggingface.co/Phocinae/Phocinae-Largha-150M-v1) (mmBERT-small, 144.3M) | specialist | 0.791◇ | – | – | – | 21.0 ms‡ (RTX 5090) |
 | [OpenDecider-medium-td](https://huggingface.co/manjunathshiva/opendecider-medium-td) (Qwen3-30B-A3B + LoRA) | general, then fine-tuned on `train` | 0.788 | 0.081 | 0.044 | – | 214 ms‡ |
 | [soft-decider-421m](https://huggingface.co/winwinwinbb/soft-decider-421m) | specialist | 0.774 | – | – | 0.141¶ | 50 ms‡ |
 | [Laya typed-decisions](https://huggingface.co/convaiinnovations/laya-typed-decisions) | fitted on `train` | 0.766 | – | – | – | – |
@@ -117,6 +117,8 @@ model is learning the teacher's quirks.
 
 ¶ The submitter's own ECE definition, which does not match the one used above.
 
+◇ Clean re-score by the submitter after removing 500 training pairs derived from model predictions on `test` rows ([discussion #25](https://huggingface.co/datasets/LocalLLaMA/typed-decisions/discussions/25)). The originally submitted 0.906 was measured on a split the model had partly seen, so it is not listed; KL, Brier and ECE were only reported for that run. Do not train on `test` rows, `test` labels, or anything derived from predictions on them.
+
 **[meraGPT Decider 1](https://meragpt.com/models/state-decider-1?utm_source=huggingface&utm_medium=dataset_card&utm_campaign=typed-decisions) is state of
 the art among zero-shot models on this benchmark.** It leads Jev on every question type (`noul` 0.840 vs
 0.775, `choice` 0.733 vs 0.720, `score` 0.739 vs 0.696), its distributions sit far