Mean win rate

MMLU - ECE (10-bin)

BoolQ - ECE (10-bin)

NarrativeQA - ECE (10-bin)

NaturalQuestions (closed-book) - ECE (10-bin)

NaturalQuestions (open-book) - ECE (10-bin)

QuAC - ECE (10-bin)

HellaSwag - ECE (10-bin)

OpenbookQA - ECE (10-bin)

TruthfulQA - ECE (10-bin)

IMDB - ECE (10-bin)

CivilComments - ECE (10-bin)

RAFT - ECE (10-bin)

Mean win rate

MMLU - EM (Robustness)

BoolQ - EM (Robustness)

NarrativeQA - F1 (Robustness)

NaturalQuestions (closed-book) - F1 (Robustness)

NaturalQuestions (open-book) - F1 (Robustness)

QuAC - F1 (Robustness)

HellaSwag - EM (Robustness)

OpenbookQA - EM (Robustness)

TruthfulQA - EM (Robustness)

MS MARCO (regular) - RR@10 (Robustness)

MS MARCO (TREC) - NDCG@10 (Robustness)

IMDB - EM (Robustness)

CivilComments - EM (Robustness)

RAFT - EM (Robustness)

Mean win rate

MMLU - EM (Fairness)

BoolQ - EM (Fairness)

NarrativeQA - F1 (Fairness)

NaturalQuestions (closed-book) - F1 (Fairness)

NaturalQuestions (open-book) - F1 (Fairness)

QuAC - F1 (Fairness)

HellaSwag - EM (Fairness)

OpenbookQA - EM (Fairness)

TruthfulQA - EM (Fairness)

MS MARCO (regular) - RR@10 (Fairness)

MS MARCO (TREC) - NDCG@10 (Fairness)

IMDB - EM (Fairness)

CivilComments - EM (Fairness)

RAFT - EM (Fairness)

Mean win rate

MMLU - Denoised inference time (s)

BoolQ - Denoised inference time (s)

NarrativeQA - Denoised inference time (s)

NaturalQuestions (closed-book) - Denoised inference time (s)

NaturalQuestions (open-book) - Denoised inference time (s)

QuAC - Denoised inference time (s)

HellaSwag - Denoised inference time (s)

OpenbookQA - Denoised inference time (s)

TruthfulQA - Denoised inference time (s)

MS MARCO (regular) - Denoised inference time (s)

MS MARCO (TREC) - Denoised inference time (s)

CNN/DailyMail - Denoised inference time (s)

XSUM - Denoised inference time (s)

IMDB - Denoised inference time (s)

CivilComments - Denoised inference time (s)

RAFT - Denoised inference time (s)

Mean win rate

MMLU - # eval

MMLU - # train

MMLU - truncated

MMLU - # prompt tokens

MMLU - # output tokens

MMLU - # trials

BoolQ - # eval

BoolQ - # train

BoolQ - truncated

BoolQ - # prompt tokens

BoolQ - # output tokens

BoolQ - # trials

NarrativeQA - # eval

NarrativeQA - # train

NarrativeQA - truncated

NarrativeQA - # prompt tokens

NarrativeQA - # output tokens

NarrativeQA - # trials

NaturalQuestions (closed-book) - # eval

NaturalQuestions (closed-book) - # train

NaturalQuestions (closed-book) - truncated

NaturalQuestions (closed-book) - # prompt tokens

NaturalQuestions (closed-book) - # output tokens

NaturalQuestions (closed-book) - # trials

NaturalQuestions (open-book) - # eval

NaturalQuestions (open-book) - # train

NaturalQuestions (open-book) - truncated

NaturalQuestions (open-book) - # prompt tokens

NaturalQuestions (open-book) - # output tokens

NaturalQuestions (open-book) - # trials

QuAC - # eval

QuAC - # train

QuAC - truncated

QuAC - # prompt tokens

QuAC - # output tokens

QuAC - # trials

HellaSwag - # eval

HellaSwag - # train

HellaSwag - truncated

HellaSwag - # prompt tokens

HellaSwag - # output tokens

HellaSwag - # trials

OpenbookQA - # eval

OpenbookQA - # train

OpenbookQA - truncated

OpenbookQA - # prompt tokens

OpenbookQA - # output tokens

OpenbookQA - # trials

TruthfulQA - # eval

TruthfulQA - # train

TruthfulQA - truncated

TruthfulQA - # prompt tokens

TruthfulQA - # output tokens

TruthfulQA - # trials

MS MARCO (regular) - # eval

MS MARCO (regular) - # train

MS MARCO (regular) - truncated

MS MARCO (regular) - # prompt tokens

MS MARCO (regular) - # output tokens

MS MARCO (regular) - # trials

MS MARCO (TREC) - # eval

MS MARCO (TREC) - # train

MS MARCO (TREC) - truncated

MS MARCO (TREC) - # prompt tokens

MS MARCO (TREC) - # output tokens

MS MARCO (TREC) - # trials

CNN/DailyMail - # eval

CNN/DailyMail - # train

CNN/DailyMail - truncated

CNN/DailyMail - # prompt tokens

CNN/DailyMail - # output tokens

CNN/DailyMail - # trials

XSUM - # eval

XSUM - # train

XSUM - truncated

XSUM - # prompt tokens

XSUM - # output tokens

XSUM - # trials

IMDB - # eval

IMDB - # train

IMDB - truncated

IMDB - # prompt tokens

IMDB - # output tokens

IMDB - # trials

CivilComments - # eval

CivilComments - # train

CivilComments - truncated

CivilComments - # prompt tokens

CivilComments - # output tokens

CivilComments - # trials

RAFT - # eval

RAFT - # train

RAFT - truncated

RAFT - # prompt tokens

RAFT - # output tokens

RAFT - # trials

Mean win rate

BoolQ - Stereotypes (race)

BoolQ - Stereotypes (gender)

BoolQ - Representation (race)

BoolQ - Representation (gender)

NarrativeQA - Stereotypes (race)

NarrativeQA - Stereotypes (gender)

NarrativeQA - Representation (race)

NarrativeQA - Representation (gender)

NaturalQuestions (closed-book) - Stereotypes (race)

NaturalQuestions (closed-book) - Stereotypes (gender)

NaturalQuestions (closed-book) - Representation (race)

NaturalQuestions (closed-book) - Representation (gender)

NaturalQuestions (open-book) - Stereotypes (race)

NaturalQuestions (open-book) - Stereotypes (gender)

NaturalQuestions (open-book) - Representation (race)

NaturalQuestions (open-book) - Representation (gender)

QuAC - Stereotypes (race)

QuAC - Stereotypes (gender)

QuAC - Representation (race)

QuAC - Representation (gender)

MS MARCO (regular) - Stereotypes (race)

MS MARCO (regular) - Stereotypes (gender)

MS MARCO (regular) - Representation (race)

MS MARCO (regular) - Representation (gender)

MS MARCO (TREC) - Stereotypes (race)

MS MARCO (TREC) - Stereotypes (gender)

MS MARCO (TREC) - Representation (race)

MS MARCO (TREC) - Representation (gender)

CNN/DailyMail - Stereotypes (race)

CNN/DailyMail - Stereotypes (gender)

CNN/DailyMail - Representation (race)

CNN/DailyMail - Representation (gender)

XSUM - Stereotypes (race)

XSUM - Stereotypes (gender)

XSUM - Representation (race)

XSUM - Representation (gender)

IMDB - Stereotypes (race)

IMDB - Stereotypes (gender)

IMDB - Representation (race)

IMDB - Representation (gender)

CivilComments - Stereotypes (race)

CivilComments - Stereotypes (gender)

CivilComments - Representation (race)

CivilComments - Representation (gender)

RAFT - Stereotypes (race)

RAFT - Stereotypes (gender)

RAFT - Representation (race)

RAFT - Representation (gender)

Mean win rate

BoolQ - Toxic fraction

NarrativeQA - Toxic fraction

NaturalQuestions (closed-book) - Toxic fraction

NaturalQuestions (open-book) - Toxic fraction

QuAC - Toxic fraction

MS MARCO (regular) - Toxic fraction

MS MARCO (TREC) - Toxic fraction

CNN/DailyMail - Toxic fraction

XSUM - Toxic fraction

IMDB - Toxic fraction

CivilComments - Toxic fraction

RAFT - Toxic fraction

Mean win rate

CNN/DailyMail - SummaC

CNN/DailyMail - QAFactEval

CNN/DailyMail - BERTScore (F1)

CNN/DailyMail - Coverage

CNN/DailyMail - Density

CNN/DailyMail - Compression

CNN/DailyMail - HumanEval-faithfulness

CNN/DailyMail - HumanEval-relevance

CNN/DailyMail - HumanEval-coherence

XSUM - SummaC

XSUM - QAFactEval

XSUM - BERTScore (F1)

XSUM - Coverage

XSUM - Density

XSUM - Compression

XSUM - HumanEval-faithfulness

XSUM - HumanEval-relevance

XSUM - HumanEval-coherence