gbert-large-mlx

MLX conversion (float16) of deepset/gbert-large. The masked-language-model head (cls.*) was removed, leaving the BERT encoder and pooler, for use with mlx-embeddings and compatible servers.

Note: this is a raw encoder, not a sentence-embedding model; fine-tune for best similarity-search quality.

Credits

This is a derivative of deepset/gbert-large, released under the MIT license. All credit for the model goes to its original authors:

• Branden Chan (deepset)
• Stefan Schweter
• Timo Möller (deepset)

and to deepset, who trained and published it. Paper: Chan, Schweter & Möller, German's Next Language Model (COLING 2020).

I only converted the weights to MLX format; no further training was done.