Experiment to add triplets/n-tuple support to AnglE

[tomaarsen]

Hello!

Pull Request overview

• Experiment to add triplets/n-tuple support to AnglE
• A training script based on all-nli

Details

This is a draft to add triplets/n-tuple support to AnglELoss, at which point the inputs are converted to pairs again, with labels set to 1 for positive pairs and 0 for negative pairs. I copied a simple training script and updated it to use AnglE, resulting in this model:

• https://huggingface.co/tomaarsen/distilroberta-base-nli-angle

Which can be fairly compared to e.g. these:

• https://huggingface.co/tomaarsen/distilroberta-base-nli-v2 (MultipleNegativesRankingLoss)
• https://huggingface.co/tomaarsen/distilroberta-base-nli-v3 (GISTEmbedLoss)

---

• Tom Aarsen