🏛️ Lucet-2B-Frontier (Roman Deity of Light & Illumination)

Lucet-2B is an autonomous reasoning, coding, and systems model built on Liquid AI's Tri-Hybrid architecture (Gated Short Convolutions + Sliding Window Attention + Global GQA) via Multi-Dimensional Smart Deep Training (MTA-SDT) on high-signal curated datasets.

Key Capabilities

1. Multi-Dimensional Training Architecture (MTA-SDT): Decoupled cognitive loss manifolds across Reasoning, Operational Action, Anti-Slop Visual/Physics Craft, and Conversational Grounding.
2. Safe Targeted SVD Uncensoring: Removes refusal directions from MLP down-projections (w2.weight) while leaving attention output projections (o_proj.weight) 100% mathematically untouched.
3. 512K (524,288) Fast ABF-RoPE: Extended context using native hardware RoPE (rope_theta = 4,000,000) without CPU YaRN interpolation overhead.
4. Calibrated GGUFs: Quantized with llama-imatrix for maximum reasoning fidelity in mohommed/Lucet-2B-Frontier-GGUF.