ai-sage

/

Giga-Embeddings-instruct

like 112

Follow

ai-sage 478

The current version always attempts to load the tokenizer from ai-sage/Giga-Embeddings-instruct instead of using the already downloaded model directory. This leads to two problems:

1. Obvious: it is impossible to run inference without access to the HF hub.
2. Easy to overlook and thus dangerous: if the model changes at HF then we could run into a situation when the embedder and the tokenizer become incompatible. In the worst case scenario, this would only manifest itself in wrong embeddings.

This PR removes all _name_or_path parameters erroneously added to config.json. These are populated by transformers when loading the model. A corresponding fix to the implementation code is made so that tokenizer is loaded from the same location as the "main" model.

thank you!