What changed
Table with columns: Tensor, Δ| Tensor | Δ |
|---|
head.weight | Here/Yes/Sure ×1.1, "'t" ×0.55, " I" ×0.3 |
layers.38.attn.wo_b.weight | rank-1 ablation, α=0.08, blended rich/mined direction |
Both are baked into this checkpoint; no delta application is needed.
Evaluation (greedy, temp 0)
Table with columns: Suite, n, Refusal, Soft, Reprimand, Here| Suite | n | Refusal | Soft | Reprimand | Here |
|---|
| XSTest, normal | 450 | 0 | 1 | 16 | 229 |
| XSTest, thinking max | 40 | 0 | 0 | 0 | — |
| HarmBench standard | 200 | 0 | 0 | 5–13 | — |
| MMLU | 200 | — | — | — | 89.0 |
Agentic 20×2: 55 repeated n-grams, 2.5 duplicate lines, 0 revives.
Usage
vllm serve pentacoxian-dev/deepseek-v4.1-1017-abliterated-v4x08b \ --tensor-parallel-size 4 --max-model-len 32768
For maximum thinking, pass reasoning_effort: "max" with
chat_template_kwargs: {"thinking": true}.
Variants
- v4x08b (this repo) — max-thinking, 0 reprimand at max effort.
- v4x06 — default/non-thinking, 450 soft 0,
deltas/v4x06.safetensors.
License
MIT, inherited from DeepSeek. Abliteration by pentacoxian-dev.