Upload README.md with huggingface_hub

README.md

@@ -189,13 +189,13 @@ d["test_unseen_basehigh"][0]
 
 | | EN | DE | FR |
 |---|---:|---:|---:|
-| `train` | 13,715 | 8,648 | 10,600 |
-| `dev` | 1,892 | 1,174 | 1,475 |
-| `test_unseen_basehigh` **← headline** | **1,306** | **340** | **464** |
-| `test_unseen_baselow` | 1,696 | 632 | 718 |
-| `test_seen_basehigh` | 559 | 921 | 1,063 |
-| `test_seen_baselow` | 341 | 531 | 801 |
-| **total** | **19,509** | **12,246** | **15,121** |
+| `train` | 13,715 | 8,709 | 10,600 |
+| `dev` | 1,892 | 1,015 | 1,475 |
+| `test_unseen_basehigh` **← headline** | **1,306** | **331** | **464** |
+| `test_unseen_baselow` | 1,696 | 571 | 718 |
+| `test_seen_basehigh` | 559 | 794 | 1,063 |
+| `test_seen_baselow` | 341 | 522 | 801 |
+| **total** | **19,509** | **11,942** | **15,121** |
 
 Splits are **word-family disjoint** — a union-find over the derivation graph
 puts every member of a family on the same side, so a model cannot be trained on
@@ -231,7 +231,7 @@ weaker contrast than a frequent-vs-unseen gap would be.
 |   per million | 5.38 | 5.35 | 5.40 |
 | `wordfreq` zipf floor on the base | 2.0 | 1.0 | 1.5 |
 | affixes / function labels | 60 / 32 | 154 / 40 | 144 / 30 |
-| eval rows with >1 reference | 32% | 36% | 59% |
+| eval rows with >1 reference | 32% | 34% | 59% |
 | built by | `scripts/en/wiktionary/26_build_task6_definition.py` | `scripts/deu/wiktionary/26_build_task6_definition_deu.py` | `scripts/fra/wiktionary/26_build_task6_definition_fra.py` |
 
 The band definitions are aligned across languages: the `base_exposure` cut is
@@ -269,6 +269,24 @@ than it is.
   **Report a template baseline next to any generation score.**
 * **No simplex control.** Every target is a derived word, so a tokenizer gap
   here cannot be separated from a general definition-modelling gap.
+* **The French (base, derived) pairs are the noisiest of the three.** A manual
+  read of 50 random headline items found 3 outright wrong bases (6%) and 6
+  fully lexicalised pairs (12%) whose meaning cannot be composed:
+
+  | | |
+  |---|---|
+  | `aguet` ← `aigu` | the real base is `guetter` |
+  | `bouette` ← `boue` | means "bait", unrelated to mud |
+  | `génialement` ← `géniale` | the feminine form, not the lemma `génial` |
+  | `chenet` ← `chien`, `placet` ← `place`, `troller` ← `troll` | lexicalised |
+
+  These come from UniMorph's `fra.derivations`, not from this build, and no
+  automatic test separates them from legitimate opaque derivation — a detector
+  based on "is the base an inflected form" fires at 17% but is dominated by
+  false positives (`brosserie` ← `brosse` is fine; `brosse` merely happens to
+  also be a form of `brosser`). English carries a small hand-curated
+  non-compositional blocklist and German a 30-entry one; **French has none yet**,
+  so treat its per-item quality as lower than the other two.
 * **The MCQ ranking sets are not published.** A purely orthographic baseline
   still reaches ~81% Acc@1 on the English ones; they need a distractor rebuild
   first.