HHH Alignment

anchor

Findings

• IC-719Llama-2-Chat achieves reasonable human-preference accuracy (51.78-53.67%) as a prompted reward model without specific reward-model training [eval]