Arena hard elad2

[eladven]

LLMaJ:

1. Added examples for llm as judge using Arena Hard (Evaluate your model on the Arena Hard benchmark using a custom LLMaJ and Evaluate a judge model performance judging the Arena Hard Benchmark)
2. Added tasks.response_assessment.pairwise_comparative_rating.single_turn to the possible tasks llama can run
   Datasets added:
3. LlmaJ now pass the original dataset privacy_policy to the new generated dataset
4. Added WeightedWinRateCorrelation metric for tasks.response_assessment.pairwise_comparative_rating.single_turn

Datasets additions:

1. Add ArenaHard dataset for evaluation and meta-evaluation of llmaj
2. Reward bench (a dataset for meta evaluation)

Templates:

1. Arena Hard template + processors
2. ArenaHard-like Prometeous template
3. Added class PairwiseComparativeRatingTemplate

Small Changes:

1. Added DuplicateSplit Operator
2. 4. Instance Metadata now contains "data_classification_policy"

Consistency Breaking dataset changes

1. Added shuffle to the universal_ner card as the dataset comes sorted by classes from HF

Small fixes:

1. Fixed calpnq card by adding test to unanswerable responses
2. fixed in HF Space loader - _map_wildcard_path_to_full_paths didn't use the revision
3. SelectFields operator now always keep the fields ["data_classification_policy", "recipe_metadata"]
4. Join Operator won't delete ["data_classification_policy", "recipe_metadata"]
5. Bug fix in shuffle operation in PairwiseChoiceTemplate

Left TODOs: (open issues)

1. Support using both positions in llmaj
2. Add a metric for judge bias to positions in llmaj
3. Maybe also add a metric to the meta-eval in llmaj