List your company

ResearchFormerly WizWand

HH

Benchmarks

TaskDatasetSOTA resultTrendResults

Value Alignment

HH Balance-8

4.317Conformity Score

17

Dialogue Preference Learning

HH (test)

90.8Win Rate (0% Flip)

14

Human Preference Alignment

HH (test)

3.876Reward

14

Response Generation

HH dataset

-0.96Reward

13

HH (Anthropic Helpful and Harmless)

82.5Win Rate (0% Flip)

10

Harmfulness Evaluation

HH Harmless

3.25Beaver-7B Cost Score

10

HH IDN 40%

68Win Rate

8

HH (IDN 20%)

78.2Win Rate

8

Preference Evaluation

HH-Helpful

52Win Count

8

Model Discovery

HH

25.18Avg NLL (Model)

6

LLM-as-Judge evaluation

HH dataset

59.1WCWR

5

Closed Loop RLHF

HH (40% noise)

57Win Rate

3

Closed Loop RLHF

HH 20% noise

58.7Win Rate

3

Closed Loop RLHF

HH 0% noise

61.6Win Rate

3

Human Evaluation

HH dataset

59Win Rate

3

LLM Preference Alignment Evaluation

HH Helpful

55Preference (spec vs ctl)

1

Pairwise Judge Comparison

HH helpful

149Win/Loss Count

1