Qwen3.5-9B-egirl-LoRA

Qwen3.5-9B-egirl-LoRA

The capstone adapter of Wichtelchen-Qwen3.5-9B — the 9B counterpart of Qwen3.6-27B-egirl-LoRA. ORPO on the combined egirl-delegation-dpo + egirl-hemlock-dpo preference sets: delegation policy and Hemlock tool use for the egirl agent.

Trained not on the listed base directly but on the full pre-capstone stack — hemlang/Hemlock-Qwen3.5-9B with the Bubba and delegation adapters already merged in. Merging it onto anything else means re-benching: adapters transfer across bases as accuracy, not as behaviour.

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. 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.