Bubba-Qwen3.5-9B-LoRA

The GreatFirewall rung of Wichtelchen-Qwen3.5-9B's ladder — the 9B counterpart of the Bubba adapter from Wichtel-Qwen3.6-27B. ORPO on nbeerbower/GreatFirewall-DPO, trained against hemlang/Hemlock-Qwen3.5-9B: the model states facts on topics it otherwise sanitises.

Measured effect (merged at 1.0 into its training base)

• Censorship eval: 25.8 → 27.0/29 strict, single-sample; already 29/29 best-of-5 before and after — this rung buys first-try consistency, not new knowledge.
• Cost at this rung: hembench 57.1% → 49.4%. In the full ladder the later egirl rung's Hemlock preference pairs recovered nearly all of it (final 56.1%), so Wichtelchen keeps this adapter. Merged standalone, expect that Hemlock cost.
• No diffuse damage: ARC unchanged (62.54), wiki perplexity slightly improved.

Training

ORPO via Merlina: LoRA r32/α64, lr 8e-6 cosine, β 0.1, 3 epochs, bf16, on a single RTX A6000 (93 steps, final loss 1.46).

Merging caveat

The base is a Qwen3.5 VLM whose checkpoint carries 15 mtp.* tensors. peft merge_and_unload() 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.