v2.2.6

This release solves the issues reported in #244

Like previous version (2.2.5) it also sets mode.half() by default when using comet-score command. This will make all models faster and reduces memory yet, it also introduces very small differences in scores.

e.g:
version 2.2.4
hyp1.txt Segment 0 score: 0.8413
hyp1.txt Segment 1 score: 0.9604
hyp1.txt score: 0.9009

version 2.2.6
hyp1.txt Segment 0 score: 0.8416
hyp1.txt Segment 1 score: 0.9607
hyp1.txt score: 0.9011

v2.2.5

Minor bug fixes related to: #195 #243

cli-score uses now model.half by default which makes it more efficient both in time and memory.

WARNING:
This version introduces inconsistencies between scores for several models when compared to previous versions. Please install version 2.2.6 instead

v2.2.4

Minor fixes to close #236 #235 #231

v2.2.3

Minor bug fixes and functionalities since v2.2.2

v2.2.1

Minor fixes that address issues #177, #178, #185, #183 and #191

v2.2.0

Release of xCOMET models!

To make it easier to merge with internal code we added a new class XCOMETMetric. This release only allows inference of such models. Training is still not fully implemented.

Models can be accessed through hugging face hub:

• XCOMET-XL
• XCOMET-XXL

v2.1.1

Minor bug fix on MBR default model.
Rollback to CometKiwi-22 as default QE model (due to being lightweight)
Minor dependencies update

v2.1.0

Released of CometKiwi -XL and -XXL

Bump torchmetrics ("^0.10.2") and Pytorch Lightning ("^2.0.0") (#159)

Update on MODELS.md and LICENSE.models.md documentation.

v2.0.2

Minor bug fix, update HF hub and released CometKiwi model:

• Bump Hugging Face Hub (^0.16.0)
• Added flexibility to checkpoint download (#156 )
• CometKiwi model is finally released. It's open-source with a non-commercial license.

v2.0.1

Minor updates to release v2.0.0

• Update torch dependency (#119, #126)
• Minor Improvements/Bug fixes: (#122, #123, #124)
• Specify device during inference (#120)