What's Changed

• Improve postprocessing efficiency by @kevinhu in #157

New Contributors

• @kevinhu made their first contribution in #157

Full Changelog: 2.1.5...2.1.6

Changelog

• Avoid unnecessary len check by using is None for tokenizer, leading to major speedups (#150)
• Change default install onnxruntime from cpu to flexible install gpu and cpu (#152)
• Allow using pre-downloaded tokenizer so SaT can be used offline (#151)
• Add checks when setting a ONNX model object (#149)

• Introduce optional hat weighting by @lsorber
• Clarify LoRA adaptation
• Clarify treat_newline_as_space: renamed to split_on_input_newlines. treat_newline_as_space will be deprecated in a future release.

• Fixes #142: AssertionError when string is only comprised of newlines, whitespace, or if its an empty strong.

• Change default behaviour for newlines in SaT.split.
  • Now, while the model ignores them, they will used to split as simple post-processing.
• Small bugfixes for LoRA training
• Update Readme for advanced usage

• Adds ONNX support for SaT models.
  • Including export scripts and an updated README.
  • This results in 50% improved inference time on GPU.

• Fix splitting of short sequences into individual characters (#127)

• Allow numpy>=2.0
• Fix adaptation code
• Add some comments

• Fixes potential CUDA device error when the input has exactly 511 tokens (#121).

• Fix a speed issue with SaT (#118). Now it is (as expected) ~6x faster than WtP.