🏛️ Lucet-2B-Frontier (Anti-Hallucination & NCM Edition)

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 Next-Concept Modeling (NCM) and R-Tuning Anti-Hallucination protocols.

Key Capabilities

1. NCM Coarse-to-Fine Planning: Structured thinking ([L1.CONCEPT] -> [L2.STRATEGY] -> [L3.COLLAPSE]), eliminating autoregressive planning errors.
2. R-Tuning "Measure Reality" Protocol: The model never guesses unobserved state; it emits deterministic measurement tools (python -c, read_file, bash).
3. Safe SVD Uncensoring: Removes refusal directions from w2.weight while leaving attention output projections (o_proj.weight) 100% mathematically untouched.
4. 512K (524,288) Fast ABF-RoPE: Extended context using native hardware RoPE (rope_theta = 4,000,000) without CPU YaRN interpolation overhead.
5. Calibrated GGUFs: Quantized with llama-imatrix for maximum reasoning fidelity in mohommed/Lucet-2B-Frontier-GGUF.