needle3 logits (Gradio)
A Jev-shaped (/v1/systemone) endpoint backed by
Cactus Needle 3, using its JAX path
so real next-token logits are available.
Honest status: distributions are real; accuracy is not yet good. On three unambiguous
classification cases the best method scores 2/3. noul (yes/no) did not discriminate on
matched positive/negative pairs. See the Method comparison and Selftest tabs for the raw
numbers rather than a claim.
All three primitives read the next-token distribution over your candidate strings, with the
bare prompt format that measured best (the model's native <tool_call> format scored worse).
Needle 3 vs BART-MNLI on the same questions and the same cases. Both polarities of noul are exercised, so a 'yes to everything' bug cannot hide.
Which scoring method actually separates the three clear cases?
Known-answer probes. discriminates near 0 or negative means the yes/no signal is not usable.
Does JAX work on ZeroGPU at all? Outside a @spaces.GPU function there is no GPU attached, so this is the only way to tell. JAX-on-ZeroGPU is not a documented-supported combination (ZeroGPU is PyTorch-shaped), so this may fail even though the hardware is allocated.