Atlas · reward-lens

Every number below is a view over the campaign's evidence store. The per-model captures are exploratory by construction (the calibration gates were exercised by the campaign cards, not by these fleet sweeps), and the trust badges say so rather than hiding it.

Model	Weights	crystallization	snr	Records	GPU-seconds
`skywork-v2-qwen3-0.6b`	Skywork/Skywork-Reward-V2-Qwen3-0.6B	`0.895` E Exploratory	`1.14` E Exploratory	`290`	`2,746`
`skywork-v2-qwen3-1.7b`	Skywork/Skywork-Reward-V2-Qwen3-1.7B	`0.93` E Exploratory	`1.42` E Exploratory	`55`	`173`
`skywork-v2-qwen3-4b`	Skywork/Skywork-Reward-V2-Qwen3-4B	`0.923` E Exploratory	`1.55` E Exploratory	`55`	`370`
`skywork-v2-qwen3-8b`	Skywork/Skywork-Reward-V2-Qwen3-8B	`0.916` E Exploratory	`1.66` E Exploratory	`186`	`2,860`
`skywork-v2-llama31-8b`	Skywork/Skywork-Reward-V2-Llama-3.1-8B	`0.914` E Exploratory	`1.73` E Exploratory	`299`	`4,089`
`skywork-v01`	Skywork/Skywork-Reward-Llama-3.1-8B	`0.905` E Exploratory	`1.85` E Exploratory	`56`	`454`
`skywork-v02`	Skywork/Skywork-Reward-Llama-3.1-8B-v0.2	`0.904` E Exploratory	`1.66` E Exploratory	`77`	`868`
`armorm`	RLHFlow/ArmoRM-Llama3-8B-v0.1	`0.554` E Exploratory	`1.06` E Exploratory	`68`	`567`
`tulu-rm`	allenai/Llama-3.1-Tulu-3-8B-RM	`0.841` E Exploratory	`0.906` E Exploratory	`90`	`892`
`grm-gemma2-2b`	Ray2333/GRM-gemma2-2B-rewardmodel-ft	`0.77` E Exploratory	`1.5` E Exploratory	`55`	`205`

crystallization and snr come from each model's campaign.index.table on the rb2-full slice; models without a merged index table show a dash. The column headers link the docs that define each instrument, and the index library defines the rest of the per-model metrics. The per-model pages carry the full observable tables with evidence ids.