Reproduce “Walrus Technical Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alkis Koudounas, Osamu Take, Hayato Futami, Quentin Jodelet, Emiru Tsunoo
Walrus uses a three-stage data curriculum to train a speech-and-audio model while keeping the Gemma 4 backbone unchanged. Task and dataset weights balance recognition with other tasks. On the CS-FLEURS Read evaluation, Walrus-R achieves a character error rate of 7.23 versus 32.30 for Gemma 4. These results support the recipe but do not isolate the effects of individual changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.