-
Release ORCA TTT-Probes (17 configurations across 3 LLMs)
Zekai Wang
17 trained Test-Time Training probes for the paper Online Reasoning
Calibration: Test-Time Training Enables Generalizable Conformal LLM
Reasoning (arXiv:2604.01170).
Each probe directory contains: probe.pt (state dict), config.json
(training hyperparameters), lambdas.json (LTT thresholds), metrics.json
(savings/error per delta), and ood_*.json (per-OOD-benchmark metrics
for Qwen2.5-32B variants).
Coverage:
- Qwen2.5-32B supervised: no_kq + qk_dh{32,64,128,256,512} + 5 architecture variants
- Qwen2.5-32B consistent: no_kq, qk_dh128
- QwQ-32B supervised: no_kq, qk_dh128
- Llama-3.3-70B supervised: no_kq, qk_dh128
-
initial commit
