Definition
Plain language
An open collection of instruction-following training data and models.
As stated in the literature
An Allen Institute open post-training dataset and model series; used as general-purpose training data whose learned evaluative rubrics transfer to specialized expert domains in EvoLM.
Why it matters: It gives researchers a freely available starting point for training capable instruction-following models without assembling their own training data from scratch.
For example, a team can take this open instruction-following collection and use it to teach a general-purpose assistant how to follow everyday requests before fine-tuning it for a narrower task.
Heard on the show
“They train the rubric generator on general-purpose Tulu data — basically nothing specialized.”Episode 019 — When the Best Reward Model Trains the Worst Policy: Inside EvoLM