ihor-matiushenko/silver-lining-gemma3-gguf

Compact on-device Small Language Model fine-tuned via Apple mlx-lm LoRA (r=16, alpha=32) from Gemma 3 270M on 1,600 multilingual empathetic cognitive reframing & gray-area safety refusal examples for the Silver Lining AI mobile app.

Quantized GGUF Artifacts

File	Size	SHA-256
silver-lining-270m-q4_k_m.gguf	249.2 MB	13e05897f2bd86b649c59e1bdf9c93f159383285fab5530fba6db435d34a9625
silver-lining-270m-q8_0.gguf	285.9 MB	644212e8f375b59f25524f833359fbfdb13c58422565a8300e3980b33670d0b6

Recommended llama.cpp / llamadart Inference Settings

• Prompt Template (parse_special = true):

  <start_of_turn>user
  [SilverLining | lang=en]
  <user_thought><end_of_turn>
  <start_of_turn>model
  

• Sampling: temp = 0.0 (or 0.1), repeat_penalty = 1.0, n_ctx = 1024, n_predict = 150.