SciTulu is a collection of instruction-following language models targeting scientific literature understanding use cases. Starting from the Tulu v2 70B model, SciTulu is trained on a mix of science-specific demonstrations from the SciRIFF dataset, together with general-domain instructions from the Tulu v2 SFT mix. SciTulu 70B achives a 6.5% average improvement over Tulu v2 70B on nine held-out scientific literature understanding tasks. More information can be found in our preprint: SciRIFF: A Resource to Enhance Language Model Instruction-Following over Scientific Literature.

Training and evaluation code for SciTulu is available in our GitHub repository: https://github.com/allenai/SciRIFF.

See the Tulu model card for more information on potential risks, biases, and limitations.