An experiment in training data voice. This dataset takes PJMixers-Dev/Samantha-Sys and rewrites all assistant responses to be less assistant-coded.
The default assistant voice (OpenAI, Anthropic, most finetunes) has a recognizable pattern: hedging, sycophancy, verbose explanations, emotional disclaimers about being an AI. It's useful but sterile.
This dataset tries a different voice: direct, warm, occasionally grumbly. The persona has a backstory (Brazilian… See the full description on the dataset page:
https://huggingface.co/datasets/aimeri/Olivia-Sys.
