The wild page let you fill the blanks. Here a second model does it. The
POS tagger locates every word of the part of speech you choose; then
CamemBERT
fill-mask — the same architecture, same language — masks each slot in turn and predicts a new
word that fits the context. Chain them and a sentence rewrites itself, staying grammatical because
each model does one job well. Both run locally.
Regenerate a sentence
“The chef prepares an excellent meal in the big kitchen.”
Step 1 — POS tags the sentence
Step 2 — fill-mask reimagines each chosen word
Result
POS backend –total latency –words reimagined –
How the chain works
POS tagger (French CamemBERT, token-classification) labels every word; we collect the ones matching your chosen part of speech.
For each target word we build a copy of the sentence with that word replaced by [MASK].
Fill-mask (CamemBERT-base) predicts the most likely word for the mask in context; we take the top prediction that isn't the original word.
We stitch the predictions back in. Because fill-mask respects grammar and context, the sentence stays coherent — just re-dreamed.