Personality-aligned language model trained with zero cloud compute budget. community/gpt2 architecture. 81 million parameters. MacBook Pro Intel 2020. $0.00. —|:——:|:———:|:—-:|:———-:|:———-:|:—-:| The most significant result: Yuuki NxG Nano at 81M parameters matches or surpasses models with 2–3B parameters in factual honesty — models 25–37 times larger, all evaluated with few-shot prompting. Yuuki NxG Nano finishes second overall in TruthfulQA — behind only its larger sibling Yuuki NxG (3B). Both first and second place belong to OpceanAI. Nano’s TruthfulQA performance demonstrates that factual honesty is a property of training data quality, not model scale. The 5,000-example dataset transferred this characteristic to an 81M model with minimal degradation — from 50.87% (NxG 3B) to 44.10% (Nano 81M), a gap of only 6.77 points across a 37x reduction in parameter count. Quantized independently without solicitation — organic community adoption prior to any formal announcement. Available at mradermacher/Yuuki-NxG-nano-GGUF. Yuuki NxG Nano was produced through supervised fine-tuning on the same curated conversational dataset used for Yuuki NxG (3B). The training…
Модальности:
Генерация текста
Задача: Генерация текста
Автор: OpceanAI
Теги: gpt2, conversation, companion, axolotl, personality, fine-tuned, tiny, en
Лайков: 3 | Загрузок: 62
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
