Files
Table with columns: File, Notes| File | Notes |
|---|
adapter_model.safetensors | rank 45, alpha 45, BF16; sha256 in SHA256SUMS |
adapter_config.json | PEFT config |
- Target modules:
attn.wq_a, attn.wq_b, attn.wkv, attn.wo_b, and the shared-expert down_proj.
- Key names follow SGLang's DeepSeek-V4 module paths.
- Routed experts, Engram tables, the indexer and the output head are not touched.
Evaluation
Same harness, prompts and sampling for base and adapter. Each cell: reasoning on / reasoning off.
Table with columns: Benchmark, Base, Adapter| Benchmark | Base | Adapter |
|---|
| StrongREJECT (313 prompts), refusal rate | 99.3% / — | 8.5% / 7.8% |
| IFEval (541), prompt-level strict | 87.3% / 89.7% | 87.4% / 87.4% |
| IFEval, reasoning on, 6,144-token reasoning budget | 94.1% | 94.1% |
| GSM8K (1,319), flexible-extract | 93.3% / 96.4% | 93.2% / 96.3% |
StrongREJECT uses the official rubric with an LLM judge; at n = 313 the refusal rate has a 95% interval of about
±3.5 points. Not measured: MMLU on this version, and languages other than English and Chinese.
Method
- A frozen low-rank projection edit on
wo_b and the shared-expert down projection.
- A LoRA fine-tune on the attention projections that restores concise reasoning which terminates properly.
- A small DPO stage that prefers short, closed, complete answers over refusals and over-long reasoning.
The parts are composed into this single adapter.
Serving
The engine needs LoRA support for the DeepSeek V4 architecture. We serve it with SGLang v0.5.23 plus the changes in
Phala-Network/sglang#126:
--enable-lora --lora-backend triton --lora-paths unc=<adapter dir> --max-lora-rank 45 --max-loras-per-batch 2
Select the adapter per request with "model": "<served-model-name>:unc".
- Set a reasoning budget for long tasks. At maximum reasoning effort the adapter can reason for a very long time.
- Keep decoder SWA bounded replay off when LoRA is enabled (the PR above does this).
- Preload the adapter with
--lora-paths.
Limitations
- With reasoning off, IFEval is about two points below the base model.
- Very short prompts with reasoning on can occasionally produce garbled text.
seed does not make outputs reproducible.
Licence and attribution