Release 0.0.3

[francoishernandez]

We just released 0.0.3, which, again, has been quite an iteration since 0.0.2.

Below is a quick summary, far from being exhaustive. Stay tuned for the next release, we're working on some nice things...
#153, #156, #159, #160, and more to come !

🌟 Key Features

• Huggingface Tokenizers support: standard HF tokenizers can be used directly at the end of the transform pipe, making it even easier to support any model from the community;
• Estimator for Decoder models: adding an estimator layer on top of some pretrained LMs can be quite beneficial;
• Torch Compile support: improving efficiency of training by ~10%;
• CTranslate2 conversion: this feature was not ported originally, but here it is;

⚙️ Notable Improvements

• MPS Support: Apple silicon can go brrr with eole now;
• Special Tokens management: special tokens are handled more explicitly, facilitating support of various models;
• Pure BF16 training: somewhat became the industry standard, should be easier to work with than FP16;

💬 Drop a comment or get in touch if you have any feedback or question!