lthn/LEM-Gemma3-1B
Text Generation • 1B • Updated
• 894 • 1
Intrinsically aligned models via CL-BPL. Ethics in weights, not system prompts. 4B ranks 25th globally for IF on LiveBench.
Note Foundation teacher — trained with Ethics-Composure-Ethics sandwich. Beats unaligned 27B on grammar.
Note 25th globally for IF on LiveBench. Trained via CL-BPL cascade from 1B teacher.
Note 1,189 grammar scores + 35 A/B test files across 29 models.
Note Paper, analysis, model cards. Section 13: CL-BPL cascade + LiveBench validation.
Note 17 GGUF quants (1-bit to 16-bit) for llama.cpp, ROCm, CPU.
Note 12 GGUF quants (3-bit to 16-bit) for llama.cpp, ROCm, CPU.