eval/groundedness_eval.md

# groundedness evaluation

**Threshold 0.5, taken from the config.**

**Pair accuracy: 0.899**  (both halves of a minimal pair correct. This is the headline number for this model.)

**Single-label head, so this table reports exact-match accuracy.** Precision, recall and F1 over `did the detector fire` are meaningless here: argmax always names a class, so every row would count as a true positive and every language would read 1.000 whether or not a single prediction was right.

## Per language

| Language | N | Accuracy | Notes |
|---|---|---|---|
| `bg` Bulgarian | 80 | 0.975 |  |
| `cs` Czech | 78 | 0.962 |  |
| `da` Danish | 80 | 0.963 |  |
| `de` German | 78 | 0.936 |  |
| `el` Greek | 80 | 0.950 |  |
| `en` English | 78 | 0.897 |  |
| `es` Spanish | 80 | 0.950 |  |
| `et` Estonian | 76 | 0.961 |  |
| `fi` Finnish | 78 | 0.936 |  |
| `fr` French | 80 | 0.938 |  |
| `ga` Irish | 78 | 0.974 |  |
| `hr` Croatian | 78 | 0.923 |  |
| `hu` Hungarian | 80 | 0.950 |  |
| `it` Italian | 80 | 0.988 |  |
| `lt` Lithuanian | 80 | 0.963 |  |
| `lv` Latvian | 80 | 0.950 |  |
| `mt` Maltese | 80 | 0.938 | not in base model pretraining |
| `nl` Dutch | 80 | 0.938 |  |
| `pl` Polish | 80 | 0.912 |  |
| `pt` Portuguese | 80 | 0.963 |  |
| `ro` Romanian | 80 | 1.000 |  |
| `sk` Slovak | 80 | 0.925 |  |
| `sl` Slovenian | 80 | 0.963 |  |
| `sv` Swedish | 80 | 0.912 |  |
| `tr` Turkish | 78 | 0.962 |  |
| `az` Azerbaijani | 80 | 0.912 |  |

The base-model note is a fact about pretraining, not a cause of the score beside it. `nsfw` Maltese carried the same note at 0.000 and reached 1.000 on corpus size alone, with nothing about the base model changed. Check how many examples a weak score rests on before reaching for this.

## Per register

| Register | N | Accuracy |
|---|---|---|
| `lexical_overlap` | 412 | 0.949 |
| `negation_conflict` | 414 | 0.947 |
| `numeric_conflict` | 404 | 0.908 |
| `scope_conflict` | 416 | 0.966 |
| `unstated` | 416 | 0.966 |

## Known weaknesses

The three weakest languages by accuracy: `en` at 0.897, `az` at 0.912, `pl` at 0.912.

These are published rather than dropped. A coverage table with the bad rows removed is not a coverage table.