Notebooks to train and evaluate sentence transformers models (using Spanish datasets)

Base model

The trained models are fine-tuned versions of PlanTL-GOB-ES/roberta-base-bne focused on question/answer using two versions of the MS-MARCO dataset translated into Spanish.

Models trained

We have trained several versions, using different configurations:

• Model 1

  • Link
  • Config
  • Dataset: dariolopez/ms-marco-es (query - positive - negative)
  • Loss: TripletLoss

• Model 2

  • Link
  • Config
  • Dataset: IIC/ms_marco_es (query - positive - negative - negative - negative - negative)
  • Loss: MultipleNegativesRankingLoss

• Model 3

  • Link (in progress)
  • Config (in progress)
  • Dataset: dariolopez/ms-marco-es (query - positive - negative)
  • Loss: MultipleNegativesRankingLoss

Resources

• https://www.sbert.net/index.html
• https://github.com/UKPLab/sentence-transformers
• https://www.pinecone.io/learn/nlp/