Ru-Jev
A multilingual cross-encoder model with 307 million parameters, trained on thirteen Russian tasks, used for typed decision tasks.
Timeline (last 90 days)
The study compares two versions of the multilingual cross-encoder Ru-Jev, of which the newer version introduces a label-wording enhancement method.
This cross-encoder has 307 million parameters and was trained on thirteen Russian tasks.
On held-out vocabularies composed of ordinary words or loanwords, the enhanced version shows higher accuracy under all settings, with improvements of up to 0.24.
Eleven of these fourteen gains passed Holm correction, the exceptions being the case of eighteen candidates for slang collocations and the case of five candidates for descriptive sentence collocations.
In the scenario where all labels are terms, the label-level confidence intervals do not contain zero under either setting.
In the abbreviation scenario, the enhanced version performs worse than the baseline on the clean slice.
For opaque code, once the labels related to the training options are removed, both versions remain at or below chance level.
All Facts (7)
The study compares two versions of the multilingual cross-encoder Ru-Jev, of which the newer version introduces a label-wording enhancement method.
50%UnverifiedThis cross-encoder has 307 million parameters and was trained on thirteen Russian tasks.
50%UnverifiedOn held-out vocabularies composed of ordinary words or loanwords, the enhanced version shows higher accuracy under all settings, with improvements of up to 0.24.
50%UnverifiedEleven of these fourteen gains passed Holm correction, the exceptions being the case of eighteen candidates for slang collocations and the case of five candidates for descriptive sentence collocations.
50%UnverifiedIn the scenario where all labels are terms, the label-level confidence intervals do not contain zero under either setting.
50%UnverifiedIn the abbreviation scenario, the enhanced version performs worse than the baseline on the clean slice.
50%UnverifiedFor opaque code, once the labels related to the training options are removed, both versions remain at or below chance level.
50%