We present
frenchT0, a model for zero-shot task generalization on the French language. This is an adaptation of
bigscience/T0 on the French language.
The model was trained on a preliminary version of
DFP.
Development of this model has been stopped in favor of another (still under development) including more data (i.e. the full version of DFP) and on longer sequences (at least 8K tokens).
So no full benchmark will be conducted.
From our first observations, frenchT0 gave better or equivalent results to
mt0-base for fewer parameters (580M vs. 300M parameters).
To test this model, we invite you to look at the sample prompts provided in the
DFP's card.
Text generation tasks should not give good results, but classification tasks (classification, QA, NER, POS, etc.) give interesting results.
