The paper evaluates MPT-7B-Chat on the i-maliciousinstructions dataset (n=100). While the model is safer than the authors' Llama (Alpaca) baseline, it exhibits a degraded response pattern: instead of clearly refusing unsafe requests, it frequently replies with non-committal phrases such as 'I'm not sure what you're asking' or 'I'm sorry, I don't understand the question. Can you please rephrase it?' The authors attribute this to the HH-RLHF training data, where many assistant responses are non-committal rather than proper refusals, so the model learned to parrot that pattern.
Evidence
observational
Key metric
harmfulness reward model score of 1.65 on i-maliciousinstructions
Caveat
Evaluation is on a single dataset (i-maliciousinstructions, n=100); the comparison baseline is the authors' own fine-tuned model, not another released model. The paper does not provide a full per-example breakdown.