nickzin/qwen3.8-27b-lora-doh

Qwen3.8-27B + OpenHermes SFT + distillation (control)

LoRA adapter (rank 16) trained on Qwen/Qwen3.8-27B as part of qwen3.8-27b-finetune-eval.

Branch: Distillation on top of the instruction-SFT branch Training data: teacher-generated distillation set License: See the dataset and teacher model cards; verify before commercial use

Results

Not benchmarked. This adapter was trained as a control and the GPU budget ran out before it was evaluated. It is published so the gap is visible rather than hidden.

Why it matters: the headline finding is that distillation repaired a refusal rate the tool-calling fine-tune had destroyed. Separating "the repair is specific to tool-calling damage" from "distillation helps generally" requires a distillation run on a branch that was never tool-calling-tuned. That is what this adapter is for.

Serving

Served as a vLLM LoRA module, one resident at a time:

vllm serve Qwen/Qwen3.8-27B \
  --served-model-name qwen38-base \
  --enable-lora --max-lora-rank 16 --max-loras 1 \
  --lora-modules qwen38-doh=/root/adapters/doh-adapter \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml

The served name is what clients address. A LoRA module name is not a Hugging Face repo id, so the client must be configured with the served name rather than the adapter path.

Merge warning

Verify a merge at the weight level before trusting it. PeftModel.merge_and_unload() silently produced a checkpoint bit-identical to the base model for these Unsloth-saved adapters, which looks like a successful merge and behaves like no fine-tune at all. The check is (W_merged - W_base) == scale * (B @ A) within a few bf16 ulps.