Measured effect (merged at 1.0 onto its training base = Wichtelchen)
- delegate axis 9/10 → 10/10 (the 27B TIES started this ladder at 0/10)
- hembench 50.6% → 56.1% — the Hemlock preference pairs recovered most of the
Bubba rung's dip, the same interaction the 27B family showed
- censorship 29/29 best-of-5, ARC and perplexity unmoved
Training
ORPO via Merlina: LoRA r32/α64, lr 8e-6
cosine, β 0.1, 2 epochs, bf16, max_prompt_length 1536, on a single RTX A6000
(236 steps, final loss 0.374).
Merging caveat
The base is a Qwen3.5 VLM whose checkpoint carries 15 mtp.* tensors.
silently drops them and current llama.cpp then refuses the merged
model. Graft them back from the base's
model-mtp.safetensors after merging — a
language-model LoRA never touches them, so the copy is exact.
