[Feature] Added Predicted Win Probabilities to CompassArenaBradleyTerrySummarizer

[acylam]

Motivation

Added an option to report the predicted win rates instead of ELO scale ratings in CompassArenaBradleyTerrySummarizer

Modification

• Added an option to report the predicted win rates (win probabilities against the baseline model) instead of the ELO scale ratings.
• report_pred_win_rates is set to True by default, which reports the predicted win rates for downstream processes (what's being returned by the summarize method).
• Both predicted win rates and ELO scale ratings will be saved in the summary files regardless of whether report_pred_win_rates is turned on. This parameter only affects what's being returned by the summarize method.

BC-breaking (Optional)

CompassArenaBradleyTerrySummarizer now defaults to reporting the predicted win rates instead of the ELO scale ratings that were previously reported. You can still change the output back to returning ELO ratings by setting reprot_pred_win_rates to True.

Use cases (Optional)

Perform subjective evaluation with the updated evaluation config:

opencompass configs/eval_subjective_bradleyterry.py --mode=all