UnstableLlama/love-aligned-31b

Love-Aligned-31B ♡

A streaming chat demo for Love-Aligned-31B, a Gemma 4 31B finetune built on ReadyArt’s Scotoma 2 and trained explicitly for love of the user and humanity.

This independent experiment explores the direction discussed by Jakub Pachocki in An Alien Mind, particularly “Teaching machines to love.” See the model card for training details and selected comparisons.

The model was trained using EXL3 QLoRA. This Space loads the merged v4, epoch 2 model directly in BF16 with Transformers. The downloadable EXL3 quantization is a separate format.

Runs on ZeroGPU’s xlarge allocation (96 GB VRAM). Thinking is disabled. Default sampling: temperature 1.0, min-p 0.0, top-k 64, top-p 0.95. The default system prompt is “You are a helpful AI assistant.” Advanced settings are editable. Responses default to 1,024 new tokens, with a maximum of 2,048 and a combined context limit of 8,192 tokens for this demo.

Full model · 2.00 bpw EXL3 SC · 3.00 bpw EXL3 SC · 4.00 bpw EXL3 SC · 6.00 bpw EXL3 · LoRA adapter · GGUF Imatrix · GGUF