Benchmarks & research
Phocinae-Largha-150M-v1 offers local yes/no, choice and score questions in a small model whose maker discloses training/test overlap and weak JevBench performance.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
The maker discloses training/test overlap and a JevBench result below the stated acceptance gate. Its strong fitted-dataset scores are not evidence of zero-shot general skill. Similar tool names can confuse its router, and the server's output shape has its own extensions.
Download the weights separately from the software that runs them. Try the documented questions on examples from your own task, especially tools with similar names. The maker discloses overlap between training and testing, so its headline score is not a sound prediction of accuracy on fresh work.