The Allen Institute for AI (AI2) has announced the release of Tülu 3, a state-of-the-art family of instruction-following models designed to set a new benchmark in AI capabilities. This release includes state-of-the-art features, methodologies, and tools, providing researchers and developers with a comprehensive, open-source solution. With Tülu 3, AI2 has successfully addressed a broad range of tasks, from conversational AI to complex problem-solving domains such as mathematics, reasoning, and evaluation.
Tülu 3 is a model family prioritizing transparency, openness, and state-of-the-art performance. The models are based on Meta’s Llama 3.1 framework and have been fine-tuned on an extensive dataset mix comprising publicly available, synthetic, and human-created data. This approach ensures that Tülu 3 achieves excellence across diverse tasks, including specialized domains like MATH, GSM8K, and IFEval while maintaining strong capabilities in general-purpose chat and reasoning tasks…
