Back in the days, WMT14 en-de dataset with 400k training samples was used a lot for NMT tasks. The reason for that is German is morphologically richer than other subsets in that benchmark.
November 25, 2024 at 10:55 AM
Back in the days, WMT14 en-de dataset with 400k training samples was used a lot for NMT tasks. The reason for that is German is morphologically richer than other subsets in that benchmark.
Have been working on a "LLM system" robustness metric "scorer".
Turns out your statistical metrics like Cohen's d and Cohen's h are really good to quantify robustness.
Cohen's h is especially good when system's output is binary.
Turns out your statistical metrics like Cohen's d and Cohen's h are really good to quantify robustness.
Cohen's h is especially good when system's output is binary.
November 25, 2024 at 10:51 AM
Have been working on a "LLM system" robustness metric "scorer".
Turns out your statistical metrics like Cohen's d and Cohen's h are really good to quantify robustness.
Cohen's h is especially good when system's output is binary.
Turns out your statistical metrics like Cohen's d and Cohen's h are really good to quantify robustness.
Cohen's h is especially good when system's output is binary.