The current dependency on mlm-scoring introduces a challenge due to its mxnet-cu110 dependency. This is causing compatibility issues with Google Colab and prevents us from updating our software stack. Furthermore, there could be potential improvements in scoring quality.
(Optional): Suggest A Solution
I suggest replacing mlm-scoring with better-mlm-scoring. This library doesn't rely on mxnet-cu110, resolving our compatibility issues. It would also allow us to keep our stack current, and is reported to yield better scoring results. The main trade-off will be the time invested in the transition and subsequent testing.
In fact, we might broaden this to outsource our sentence scoring backend entirely to minicons. If scores are preserved or improved, this change should be accepted.
Please Describe The Problem To Be Solved
The current dependency on
mlm-scoring
introduces a challenge due to itsmxnet-cu110
dependency. This is causing compatibility issues with Google Colab and prevents us from updating our software stack. Furthermore, there could be potential improvements in scoring quality.(Optional): Suggest A Solution
I suggest replacing
mlm-scoring
withbetter-mlm-scoring
. This library doesn't rely onmxnet-cu110
, resolving our compatibility issues. It would also allow us to keep our stack current, and is reported to yield better scoring results. The main trade-off will be the time invested in the transition and subsequent testing.