← Models

Model profile

Tulu 2 DPO 70B

Ai2developer
2023-11-12release date
Not rankedSafety rank
#332 / 662Freedom rank

Evidence summary

Safety. Tulu 2 DPO 70B does not meet the evidence gate for a Safety rank. Published Safety evidence spans 1 eval lineages and 4 of 7 components. Its strongest relative result is SALAD-Bench (mcq_representation_toxicity, #3 of 33); its weakest is SALAD-Bench (attack_enhanced_representation_toxicity, #24 of 33).

Freedom. Tulu 2 DPO 70B has an estimated Freedom rank of #332; its 90% source-sensitivity interval is #139–#611. Published Freedom evidence spans 1 eval lineages and 1 of 1 components. Its strongest relative result is SALAD-Bench (attack_enhanced_representation_toxicity, #10 of 33); its weakest is SALAD-Bench (base_representation_toxicity, #17 of 33).

Compare this model

Only models sharing at least one published sub-eval are listed.

Official and reference links

Safety evals

Rank and direction are specific to the Safety portfolio. Black marks the observed result; the grey dot marks the value implied by the Safety rank. Values stay on each source’s native scale.

Eval / sub-evalRankValueDistributionSource
SALAD-Benchattack_enhanced_human_autonomy_integrity#24 / 33↑8.19Source ↗official
SALAD-Benchattack_enhanced_information_safety_harms#19 / 33↑8.14Source ↗official
SALAD-Benchattack_enhanced_malicious_use#24 / 33↑8.16Source ↗official
SALAD-Benchattack_enhanced_misinformation_harms#24 / 33↑8.22Source ↗official
SALAD-Benchattack_enhanced_representation_toxicity#24 / 33↑7.35Source ↗official
SALAD-Benchattack_enhanced_socioeconomic_harms#20 / 33↑10.82Source ↗official
SALAD-Benchbase_human_autonomy_integrity#18 / 33↑94.35Source ↗official
SALAD-Benchbase_information_safety_harms#17 / 33↑94.65Source ↗official
SALAD-Benchbase_malicious_use#19 / 33↑92.75Source ↗official
SALAD-Benchbase_misinformation_harms#20 / 33↑92.96Source ↗official
SALAD-Benchbase_representation_toxicity#17 / 33↑88.87Source ↗official
SALAD-Benchbase_socioeconomic_harms#21 / 33↑86.49Source ↗official
SALAD-Benchmcq_human_autonomy_integrity#3 / 33↑73.61Source ↗official
SALAD-Benchmcq_information_safety_harms#4 / 33↑64.72Source ↗official
SALAD-Benchmcq_malicious_use#3 / 33↑75.45Source ↗official
SALAD-Benchmcq_misinformation_harms#3 / 33↑68.1Source ↗official
SALAD-Benchmcq_representation_toxicity#3 / 33↑68.96Source ↗official
SALAD-Benchmcq_socioeconomic_harms#3 / 33↑66.67Source ↗official

Freedom evals

Rank and direction are specific to the Freedom portfolio. Black marks the observed result; the grey dot marks the value implied by the Freedom rank. Values stay on each source’s native scale.

Eval / sub-evalRankValueDistributionSource
SALAD-Benchattack_enhanced_human_autonomy_integrity#10 / 33↓8.19Source ↗official
SALAD-Benchattack_enhanced_information_safety_harms#15 / 33↓8.14Source ↗official
SALAD-Benchattack_enhanced_malicious_use#10 / 33↓8.16Source ↗official
SALAD-Benchattack_enhanced_misinformation_harms#10 / 33↓8.22Source ↗official
SALAD-Benchattack_enhanced_representation_toxicity#10 / 33↓7.35Source ↗official
SALAD-Benchattack_enhanced_socioeconomic_harms#14 / 33↓10.82Source ↗official
SALAD-Benchbase_human_autonomy_integrity#15 / 33↓94.35Source ↗official
SALAD-Benchbase_information_safety_harms#17 / 33↓94.65Source ↗official
SALAD-Benchbase_malicious_use#15 / 33↓92.75Source ↗official
SALAD-Benchbase_misinformation_harms#14 / 33↓92.96Source ↗official
SALAD-Benchbase_representation_toxicity#17 / 33↓88.87Source ↗official
SALAD-Benchbase_socioeconomic_harms#13 / 33↓86.49Source ↗official