Antoine Chaffin

Adding a list of the instruction prompts used during training

2

Hello, I am trying to test the outputs of OFA (by hand) and I find it a bit difficult to be 100% sure of how to prompt the model to...

Issue with indexing BGE-M3 (large dimensionality vectors)

8

Hello, I am trying to use the [BGE-M3](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/BGE_M3) model within the library. I managed to tweak the library a bit to be able to use the model to do the...

Set default to non normalization of the scores

1

Following JaColBERTv2.5, I added the [normalization of the scores for distillation.](https://github.com/lightonai/pylate/blob/d526d98b1b44899c2ba7e2c20addb22be2faecbd/pylate/losses/distillation.py#L112) Yet, this setup is not [really standard/well-know](https://github.com/lightonai/pylate/discussions/58) and so the default should be no normalization, with a proper documentation...

Right now, the pooling method pools the embedding tokens to a fixed ratio. We should allow users to set a threshold on the cosine similarity up to which the tokens...

Only pad documents to largest documents in the column

1

Right now, for simplicity, during distillation we pad every document to the max length so we can easily stack them to compute the scores. An optimization would be to only...

Leverage existing projection layers from ST models

1

Right now, when initializing from a ST checkpoints, we chop-off the eventual "Dense" module. Although these checkpoints require training anyways, this layer can be a good initialization for the linear...

Right now, the normalization is done "by hand" in the encode function or need to be done afterwards when the forward pass is used. It would be better to use...

@tomaarsen pointed out that the logging of step/epoch of the triplet evaluator in the model card might be broken, pointing [this model card as an example](https://huggingface.co/yosefw/colbert-roberta-amharic-medium#training-logs). I did not have...

As explained in #91, query expansion do not "work" when using flash attention, unattended tokens return embeddings full of zeros. One issue I did not see coming is that a...

Call create_collection during the first add_document call in Voyager index to infer embedding dimension

1

Right now, we are creating the Voyager index during the initialization of the index model, which force us to have the target embedding dimension and so to ask the user...