Schierling-Qwen3.5-9B-LoRA

Schierling-Qwen3.5-9B-LoRA

The Hemlock language adapter for Qwen3.5-9B — the 9B counterpart of Schierling-Qwen3.6-27B-LoRA. Supervised fine-tuning on hemlang/Hemlock-SFT-combined (docs and semantics, execution-verified generation, and C/Python→Hemlock translation, decontaminated against hembench).

Results

Measured on hembench with every generated program executed by the real interpreter (Q8_0, greedy). The comparison point is the sibling ...-TIES-Hemlock-SFT model (the same base with a lighter Hemlock SFT); "after" is the plain TIES base with this adapter merged:

No measurable general-capability cost: ARC 61.5→62.5, wiki perplexity +1.9%. As with the 27B, stdlib is the weak axis — exact function names and signatures are memorised API knowledge; check calls against the docs.

Training

SFT via Merlina: LoRA r32/α64, dropout 0.05, lr 2e-4 cosine, 2 epochs, bf16, max_length 2048, on a single RTX A6000 (718 steps, final loss 0.090).

Merging caveat

The base is a Qwen3.5 VLM whose checkpoint carries 15 mtp.* tensors (the multi-token- prediction head). peft merge_and_unload() silently drops them, and current llama.cpp then refuses to load the merged model. Graft them back from the base's model-vlm.safetensors after merging — they are untouched by a language-model LoRA, so the copy is exact.

Note on the base

Huihui-Qwen3.5-9B-abliterated-TIES also has a sibling with a lighter Hemlock SFT already baked in (...-TIES-Hemlock-SFT, trained on the 2,673-row base set only — 28.5% hembench). This adapter was trained on the plain TIES base and supersedes that rung.