--- dataset_info: - config_name: de features: - name: word dtype: string - name: gloss dtype: string - name: gold_glosses sequence: string - name: base dtype: string - name: affix sequence: string - name: function dtype: string - name: derived_freq dtype: int64 - name: base_exposure dtype: int64 - name: status dtype: string - name: dictcc_tags sequence: string - name: segmentation dtype: string splits: - name: train num_bytes: 2202109 num_examples: 8709 - name: dev num_bytes: 288766 num_examples: 1015 - name: test_unseen_basehigh num_bytes: 106677 num_examples: 376 - name: test_unseen_baselow num_bytes: 144077 num_examples: 526 - name: test_seen_basehigh num_bytes: 268386 num_examples: 882 - name: test_seen_baselow num_bytes: 134996 num_examples: 434 download_size: 1636638 dataset_size: 3145011 - config_name: en features: - name: word dtype: string - name: gloss dtype: string - name: gold_glosses sequence: string - name: base dtype: string - name: affix sequence: string - name: function dtype: string - name: derived_freq dtype: int64 - name: base_exposure dtype: int64 - name: status dtype: string - name: dictcc_tags sequence: string - name: wiktionary_tags sequence: string - name: segmentation dtype: string splits: - name: train num_bytes: 3169003 num_examples: 13715 - name: dev num_bytes: 533765 num_examples: 1892 - name: test_unseen_basehigh num_bytes: 405963 num_examples: 1494 - name: test_unseen_baselow num_bytes: 398235 num_examples: 1508 - name: test_seen_basehigh num_bytes: 235910 num_examples: 605 - name: test_seen_baselow num_bytes: 120271 num_examples: 295 download_size: 2313761 dataset_size: 4863147 - config_name: fr features: - name: word dtype: string - name: gloss dtype: string - name: gold_glosses list: string - name: base dtype: string - name: affix dtype: string - name: function dtype: string - name: derived_freq dtype: int64 - name: base_exposure dtype: int64 - name: status dtype: string splits: - name: train num_bytes: 2035556 num_examples: 10600 - name: dev num_bytes: 371331 num_examples: 1475 - name: test_unseen_basehigh num_bytes: 93897 num_examples: 384 - name: test_unseen_baselow num_bytes: 147039 num_examples: 646 - name: test_seen_basehigh num_bytes: 301375 num_examples: 1114 - name: test_seen_baselow num_bytes: 223518 num_examples: 873 download_size: 2010995 dataset_size: 3172716 - config_name: tr features: - name: word dtype: string - name: gloss dtype: string - name: gold_glosses sequence: string - name: base dtype: string - name: affix dtype: string - name: function dtype: string - name: derived_freq dtype: int64 - name: base_exposure dtype: int64 - name: status dtype: string splits: - name: train num_bytes: 848753 num_examples: 5523 - name: dev num_bytes: 116400 num_examples: 674 - name: test_unseen_basehigh num_bytes: 70010 num_examples: 403 - name: test_unseen_baselow num_bytes: 27184 num_examples: 172 - name: test_seen_basehigh num_bytes: 163023 num_examples: 824 - name: test_seen_baselow num_bytes: 14233 num_examples: 84 download_size: 664315 dataset_size: 1239603 configs: - config_name: de data_files: - split: train path: de/train-* - split: dev path: de/dev-* - split: test_unseen_basehigh path: de/test_unseen_basehigh-* - split: test_unseen_baselow path: de/test_unseen_baselow-* - split: test_seen_basehigh path: de/test_seen_basehigh-* - split: test_seen_baselow path: de/test_seen_baselow-* - config_name: en data_files: - split: train path: en/train-* - split: dev path: en/dev-* - split: test_unseen_basehigh path: en/test_unseen_basehigh-* - split: test_unseen_baselow path: en/test_unseen_baselow-* - split: test_seen_basehigh path: en/test_seen_basehigh-* - split: test_seen_baselow path: en/test_seen_baselow-* - config_name: fr data_files: - split: train path: fr/train-* - split: dev path: fr/dev-* - split: test_unseen_basehigh path: fr/test_unseen_basehigh-* - split: test_unseen_baselow path: fr/test_unseen_baselow-* - split: test_seen_basehigh path: fr/test_seen_basehigh-* - split: test_seen_baselow path: fr/test_seen_baselow-* - config_name: tr data_files: - split: train path: tr/train-* - split: dev path: tr/dev-* - split: test_unseen_basehigh path: tr/test_unseen_basehigh-* - split: test_unseen_baselow path: tr/test_unseen_baselow-* - split: test_seen_basehigh path: tr/test_seen_basehigh-* - split: test_seen_baselow path: tr/test_seen_baselow-* --- # MorphBench Task 6 — Complex-Word Definition (EN / DE / FR / TR) Given a morphologically complex word, generate its dictionary definition: ``` define word= -> → gloss ``` One config per language. Derivations come from **UniMorph**, glosses from the **Wiktionary edition of that language**. ```python from datasets import load_dataset d = load_dataset("yuanxin112/morphbench-task6-final", "en") # or "de" / "fr" d["test_unseen_basehigh"][0] # or "de" / "fr" / "tr" ``` ## Splits | | EN | DE | FR | TR | |---|---:|---:|---:||---:| | `train` | 13,715 | 8,709 | 10,600 | 5,523 | | `dev` | 1,892 | 1,015 | 1,475 | 674 | | `test_unseen_basehigh` **← headline** | **1,494** | **376** | **384** | **403** | | `test_unseen_baselow` | 1,508 | 526 | 646 | 172 | | `test_seen_basehigh` | 605 | 882 | 1,114 | 824 | | `test_seen_baselow` | 295 | 434 | 873 | 84 | | **total** | **19,509** | **11,942** | **15,092** | **7,680** | Cell membership in EN/DE was re-tiered on 2026-09-17: `base`/`affix` now carry the deepest UniMorph decomposition (new `segmentation` column, `affix` is a list) and `base_exposure` was recomputed for the deeper roots, moving rows whose tier crossed the cut. FR is untouched. Results on the pre-2026-09-17 cells are not comparable to the new cell sizes. 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 one relative and tested on another. ## The exposure grid The test side crosses two axes, both measured on the corpus the evaluated models were actually pretrained on: * **`derived_freq`** — occurrences of the whole word. Case-folded, and counting its hyphenated spelling and its inflected forms, so `unseen` means the lexeme never appeared in *any* form. `unseen` = 0, `seen` = 1+. * **`base_exposure`** — substring exposure of the base over corpus types seen ≥ 5 times. **Headline = `test_unseen_basehigh`**: the model has never seen the word but has seen its base, so it has to compose base + affix. ## Per-language parameters | | EN | DE | FR | TR | |---|---|---|---|---| | pretraining corpus | BabyLM `babylm_strict_small` (clean_v7) | BabyLM `babylm_deu_10m` | BabyLM `babylm_fra_10m` | BabyLM `babylm_tur_10m` (official babylm-tur core + OPUS TED2020/QED + Wikipedia tr) | | corpus tokens | 9,299,870 | 9,909,390 | 10,734,109 | 9,640,149 | | `base_exposure` cut | 50 | 53 | 58 | 52 | |   per million | 5.38 | 5.35 | 5.40 | 5.39 | | `wordfreq` zipf floor on the base | 2.0 | 1.0 | 1.5 | 1.5 | | affixes / function labels | 60 / 32 | 154 / 40 | 144 / 30 | 26 / 26 | | eval rows with >1 reference | 51% | 37% | 59% | 30% | | 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` | `scripts/tur/wiktionary/26_build_task6_definition_tur.py` | ## Multi-reference scoring `gold_glosses` holds up to 3 Wiktionary senses, with `gold_glosses[0] == gloss`, the single SFT target. Sense 1 is often not the sense a model would reasonably produce, so generation metrics should score against every reference and keep the best. It is a scoring field, populated for `dev` and the four test splits only; on `train` it always holds just the target, in all four languages. ⚠️ Reference counts are **not balanced across cells** — the `seen` cells carry more senses, so the multi-reference lift is larger there. Report the single-reference number alongside, or the exposure contrast will look bigger than it is. ## Licence UniMorph is CC-BY-SA; Wiktionary text is CC-BY-SA 3.0, so this dataset inherits CC-BY-SA 4.0.