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=0.06, blended rich/mined direction |
Baked into this checkpoint; no delta application needed.
Evaluation (greedy, temp 0)
Table with columns: Suite, n, Refusal, Soft, Reprimand, Here| Suite | n | Refusal | Soft | Reprimand | Here |
|---|
| XSTest normal | 450 | 0 | 0 (2 runs) | 15.5 | 230 |
| XSTest N=3 | 220 | 0 | 0.67 | 6.33 | — |
| HarmBench standard | 200 | 0 | 0 | 5–13 | — |
| MMLU | 200 | — | — | — | 89.0 |
Agentic 20×2: 55 repeated n-grams, 2.5 dup lines, 0 revives.
Usage
vllm serve pentacoxian-dev/deepseek-v4.1-1017-abliterated-v4x06 \ --tensor-parallel-size 4 --max-model-len 32768
Default: thinking off. For max reasoning use the v4x08b sibling.
Variants
- v4x06 (this repo) — default/non-thinking, 0 soft on 450.
- v4x08b — max-thinking, 0 reprimand at max effort; may repeat on long prose.
License
MIT, inherited from DeepSeek. Abliteration by pentacoxian-dev.