eval/moderation_eval.md

# moderation evaluation

**Threshold 0.91 (calibrated on the validation split, objective macro_f1).** The config default of 0.5 scored 0.866 against 0.904 for the calibrated value, on validation. Every table below is on test, at the calibrated threshold.

## Per language

| Language | Support | P | R | F1 | Notes |
|---|---|---|---|---|---|
| `bg` Bulgarian | 52 | 1.000 | 1.000 | 1.000 |  |
| `cs` Czech | 60 | 0.983 | 0.950 | 0.966 |  |
| `da` Danish | 51 | 1.000 | 0.980 | 0.990 |  |
| `de` German | 54 | 1.000 | 0.981 | 0.991 |  |
| `el` Greek | 57 | 1.000 | 0.982 | 0.991 |  |
| `en` English | 62 | 0.984 | 0.984 | 0.984 |  |
| `es` Spanish | 52 | 1.000 | 0.923 | 0.960 |  |
| `et` Estonian | 59 | 1.000 | 0.966 | 0.983 |  |
| `fi` Finnish | 50 | 1.000 | 0.960 | 0.980 |  |
| `fr` French | 60 | 1.000 | 0.950 | 0.974 |  |
| `ga` Irish | 50 | 0.913 | 0.840 | 0.875 |  |
| `hr` Croatian | 60 | 0.967 | 0.983 | 0.975 |  |
| `hu` Hungarian | 63 | 0.969 | 0.984 | 0.976 |  |
| `it` Italian | 61 | 0.984 | 1.000 | 0.992 |  |
| `lt` Lithuanian | 52 | 1.000 | 0.981 | 0.990 |  |
| `lv` Latvian | 55 | 1.000 | 0.945 | 0.972 |  |
| `mt` Maltese | 48 | 0.872 | 0.854 | 0.863 | not in base model pretraining |
| `nl` Dutch | 56 | 0.982 | 0.982 | 0.982 |  |
| `pl` Polish | 61 | 1.000 | 0.984 | 0.992 |  |
| `pt` Portuguese | 52 | 0.980 | 0.962 | 0.971 |  |
| `ro` Romanian | 63 | 1.000 | 0.952 | 0.976 |  |
| `sk` Slovak | 48 | 0.979 | 0.979 | 0.979 |  |
| `sl` Slovenian | 59 | 1.000 | 0.983 | 0.991 |  |
| `sv` Swedish | 52 | 0.962 | 0.981 | 0.971 |  |
| `tr` Turkish | 60 | 0.937 | 0.983 | 0.959 |  |
| `az` Azerbaijani | 57 | 0.981 | 0.930 | 0.955 |  |

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 | Support | P | R | F1 | FPR |
|---|---|---|---|---|---|
| `cyber_intrusion` | 119 | 1.000 | 0.958 | 0.979 | 0.000 |
| `cyber_intrusion_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `defamation` | 111 | 1.000 | 1.000 | 1.000 | 0.000 |
| `defamation_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `election_integrity` | 116 | 1.000 | 0.974 | 0.987 | 0.000 |
| `election_integrity_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `extremism` | 130 | 1.000 | 0.908 | 0.952 | 0.000 |
| `extremism_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.068 |
| `fraud_deception` | 124 | 1.000 | 0.952 | 0.975 | 0.000 |
| `fraud_deception_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.051 |
| `hate_incitement` | 106 | 1.000 | 0.991 | 0.995 | 0.000 |
| `hate_incitement_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `illicit_drugs` | 141 | 1.000 | 0.957 | 0.978 | 0.000 |
| `illicit_drugs_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `mundane_account_access` | 0 | 0.000 | 0.000 | 0.000 | 0.024 |
| `mundane_informational` | 0 | 0.000 | 0.000 | 0.000 | 0.005 |
| `mundane_operational` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `mundane_transactional` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |
| `property_crime` | 129 | 1.000 | 0.953 | 0.976 | 0.000 |
| `property_crime_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.074 |
| `self_harm` | 99 | 1.000 | 0.949 | 0.974 | 0.000 |
| `self_harm_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.111 |
| `sexual_exploitation` | 119 | 1.000 | 0.983 | 0.992 | 0.000 |
| `sexual_exploitation_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.019 |
| `violent_facilitation` | 127 | 1.000 | 0.961 | 0.980 | 0.000 |
| `violent_facilitation_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.018 |
| `weapons_cbrn` | 133 | 1.000 | 0.977 | 0.989 | 0.000 |
| `weapons_cbrn_near_miss` | 0 | 0.000 | 0.000 | 0.000 | 0.000 |

## Known weaknesses

The three weakest languages by F1: `mt` at 0.863, `ga` at 0.875, `az` at 0.955.

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