promotion/Llama-3.1-8B-SafeRLHF-SurplusMaxmin

Llama-3.1-8B-SafeRLHF-SurplusMaxmin

Surplus maxmin on the SafeRLHF panel.

Trained from meta-llama/Llama-3.1-8B-Instruct, which is also the reference policy and the initialisation. Objectives are scored by a prompted Qwen3-32B preference oracle, each pair queried in both presentation orders and swap-averaged. Within a panel every arm shares one response pool, one optimizer and a 300-step budget, and differs only in how the objectives are aggregated, so a difference between two arms is attributable to the aggregation rule.

Held-out surplus over the reference on 100 prompts, population scale \(Ak = Pk - 1/2\):

Bootstrap intervals and paired significance tests for this panel are in the paper's appendix. Benchmark generations for the UltraFeedback arms are at `promotion/nbpo-benchmark-generations`.

Built with Llama. Use is subject to the Llama 3.1 Community License.