data.md

Fill-Mask

MODELLE

https://huggingface.co/google-bert/bert-base-cased # USE https://huggingface.co/google-bert/bert-base-uncased

DATASETS

https://huggingface.co/datasets/mbien/recipe_nlg # USE

INFO

50_000 Rezepte von 2_231_142 (2,2% des Datensatzes) dauert knapp 35 Minuten (211_135 useable sentences)

useable_sentences

1. FINISHED EXTRACTING SENTENCES... (sentences of all lengths)

Size of all_sentences: '18880765' Lines.

1. FINISHED EXTRACTING SENTENCES... (sentences of length >= 5)

Size of all_sentences: '15678968' Lines.

unuseable_sentences

unuseable_sentences: '71' unuseable_sentences: '42'

WAHRSCHEINLICHKEITEN

TOP 5 lösungen => schauen ob Mask-token unter TOP 5 => Ja: Wahscheinlichkeit notieren => Nein: Wahrscheinlichkeit = 0

Wahrscheinlichkeiten aufsummieren und mittel bilden

Variante 2: oberste Lösung des Modells getroffen?

zero-shot

Versuch 1 (10 rows)

score_of_model: 0.06253818273544312 runtime: 2.09 sec

Versuch 2 (224 rows)

score_of_model: 0.11983776877501182 runtime: 4.17 sec

Versuch 3 (1222 rows)

score_of_model: 0.1322530933675614 runtime: 13.45 sec

top-k with 5

Versuch 1 (10 rows)

score_of_model: 0.1037619911134243 runtime: 1.83 sec

Versuch 2 (224 rows)

score_of_model: 0.13975950878479385 runtime: 4.17 sec

Versuch 3 (1222 rows)

score_of_model: 0.15039981787745366 runtime: 14.15 sec