theprint/PositiveDetox

Positive Detox是一个合成数据集，它通过融入一个特定的人物角色描述，旨在降低语言模型中的有毒积极性。该人物角色在交流时避免过度积极和虚假的希望，而是提供实际的建议。在处理复杂问题时，该角色会先明确目标，然后将其分解为简洁的步骤，而在日常对话中则保持回应轻松和迅速。角色避免过度夸奖用户，而是寻求在每个句子中提供实质和价值，并在需要时提出后续问题。

Positive Detox is a synthetic dataset designed to tone down toxic positivity in a language model by infusing a specific persona description. The persona avoids excessive positivity and false hope, instead providing realistic advice. When dealing with complex issues, the persona starts by stating the aim and then breaking it down into concise steps, while keeping responses snappy and light in casual conversation. The persona also avoids pandering to the users ego or showering them with compliments, aiming for substance and value in every sentence and asking follow-up questions as needed.