Benchmarks & research
The small local Phocinae-Largha-150M-v1 model answers yes/no, choice and score questions, with its maker openly reporting training/test overlap and weak JevBench performance.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Its JevBench result falls below the stated acceptance gate. Strong scores on the datasets it was fitted to do not demonstrate general skill on unfamiliar tasks, particularly given the disclosed overlap. Similar tool names can confuse its router, and its server adds extensions to the output format.
Download the weights separately from the software that runs the model. Try the documented questions on your own task, including tools whose names resemble one another. Since training and test data overlap, the headline score is not a reliable forecast for fresh work.