AI2 releases Tulu 3 post-training recipe

The Allen Institute for AI (AI2) released Tülu 3, an open post-training recipe covering 8-billion and 70-billion-parameter models along with intermediate training checkpoints, complete datasets, and the full open-instruct codebase used to build them. AI2 framed the release as addressing a specific gap: while pretrained open-weight base models such as Llama had become widely available, the post-training methods labs used to turn a base model into a capable, well-behaved assistant remained largely undisclosed, even by developers of otherwise open models. AI2 noted that not one of the top 50 models on the LMSYS Chatbot Arena leaderboard had published its post-training data.

Tülu 3’s recipe combined several stages: large-scale synthetic instruction data generation, preference data built from the model’s own on-policy generations rather than external comparisons, and reinforcement learning with verifiable rewards, a technique that trained specific skills such as maths and coding against rewards that could be checked automatically for correctness rather than requiring a learned reward model. AI2 released the datasets covering knowledge recall, instruction-following, reasoning, maths, coding and multilingual capability, decontaminated against public benchmarks, alongside detailed documentation of the training procedure and evaluation methodology.

By publishing every component of the pipeline rather than only final weights, Tülu 3 was positioned as a reproducibility benchmark for open post-training research, aimed at narrowing the gap between what outside researchers could study and what frontier labs like OpenAI and Anthropic kept proprietary. AI2 followed it days later with OLMo 2, a fully open base model built to pair with the Tülu 3 recipe.