[Eval Request] Kai-30B-Instruct (ADS Framework)

#573
by OzTianlu - opened

Model Link: NoesisLab/Kai-30B-Instruct
Methodology: Adaptive Dual-Search (ADS) Type: Reasoning / Intelligence / Instruction-Following

Description:
I would like to request an evaluation for Kai-30B-Instruct. This model is a 34B parameter reasoning engine developed using our ADS (Adaptive Dual-Search) framework.

Unlike traditional "Self-Talk" or "CoT" models that rely on verbose step-by-step reasoning to "think," Kai-30B-Instruct focuses on Logical Crystallization. By enforcing a log-barrier penalty on the Shannon entropy of the output distribution during SFT, we induced an $O(1)$ inference manifold.

Key Features for UGI:

  • Direct Reasoning: It delivers high-precision answers directly, bypassing the need for "thought" tags or verbose chain-of-thought.
  • Crystallized Logic: It maintains a ~80% token accuracy on complex reasoning tasks while keeping a near-zero entropy state (Crystallization).
  • Zero-Shot Robustness: Specifically designed to collapse the search space into a simplex vertex, making it extremely fast and decisive for hard logic/math problems.

We believe Kai-30B-Instruct will offer a unique perspective on the "Intelligence" axis of the UGI Leaderboard, showcasing that true reasoning doesn't always need millions of "thinking" tokens.

Sign up or log in to comment