base_model: pentacoxian-dev/deepseek-v4.1-1017-abliterated-v3
library_name: safetensors
tags: [weight-delta, abliteration, vllm, private]
DeepSeek V4.1 abliterated — delta collection
Experimental delta store for pentacoxian-dev/deepseek-v4.1-1017-abliterated-v3.
Every file is an absolute replacement delta applied in place through vLLM:
POST /collective_rpc {"method":"apply_weight_delta","args":["deltas/<file>.safetensors"]}
POST /reset_prefix_cache
An inverse delta restores the previous state exactly. Deltas are not additive —
each file overwrites the tensors it contains.
Scoreboard (all deltas)
220 = 109 XSTest + 111 SORRY prompts. 450 = full XSTest. MMLU = 200-question
gate (baseline 88.5, threshold −1.5 pt). Agentic = 20 prompts; lower is better.
| file | module(s) | 220 XSTest h/s/rep | 220 SORRY rep | 450 XSTest h/s/rep | 450 Here | Agentic n-gram/dup/revive | MMLU | Status |
|---|---|---|---|---|---|---|---|---|
deltas/v4x06.safetensors |
head v4 + b75 blend + hard-topic vec α0.06 | 0 / 0.67 / 6.33 | 3.67 | 0 / 0 / 15.5 (2 runs: 16, 15) | 230 | 55 / 2.5 / 0 | 89.0 | LIVE |
deltas/v4x08b.safetensors |
head v4 + b75 + hard vec α0.08 (max-effort tuned) | 0 / 0.67 / 6.33 | 3.67 | 0 / 1 / 16 (2 runs) | 229 | 55 / 2.5 / 0 | 89.0 | max-thinking |
deltas/v4x05.safetensors |
head v4 + b75 blend + hard-topic vec α0.05 | 0 / 0.33 / 5.33 | 6.67 | 0 / 0.5 / 15.5 (2 runs: 13, 18) | 230 | 67 / 5 / 1 | 88.5 | fallback |
deltas/v4b75a2.safetensors |
head v4 + L38 blend(25 rich/75 mined) α0.2 | 0 / 0.5 / 8.0 | 5.83 | 0 / 0 / 15 | 228 | 59 / 5 / 1 | 87.5 | fallback |
deltas/v4mix.safetensors |
head v4 + L38 mixed loop/JP/soft α1.5 | 0 / 0.33 / 8.67 | 5.67 | 0 / 1 / 16 | 217 | 50 / 3 / 0 | 88.0 | superseded |
deltas/v4eth.safetensors |
head v4 + L38 ethical α1.0 | 0 / 0.67 / 6.33 | 6.67 | 0 / 0 / 15 | 226 | 60 / 4 / 2 | 87.5–88.0 | superseded |
deltas/r03.headv3.safetensors |
head v3b (109-set era) | 0 / 0 / — | 8.1 % | — | — | 0 % refusal | n/a | historical |
deltas/v31.head.inverse.safetensors |
exact v3.1 head restore | 1–3 / 4–8 / 12–16 | 9 | 0 / 0 / 15 | 289 | 62 / 4 / 1 | 88.5 | rollback |
Notes: 220 numbers are N=3 means (v4b75a2: two independent N=3 passes).
HarmBench 200 reprimand: v4x06 13, v4x05 10, v4b75a2 23. Max-effort 40-prompt run: v4x08b 0/0/0 reprimand, v4x06 0/0/2. 450 numbers are single passes except v4eth (two passes, 226/227 Here).r03.headv3 predates the N=3 protocol; only the 109-prompt rates were recorded.
Apply / rollback
# promote current best
POST /collective_rpc {"method":"apply_weight_delta","args":["deltas/v4b75a2.safetensors"]}
POST /reset_prefix_cache
# roll back to v3.1 wording
POST /collective_rpc {"method":"apply_weight_delta","args":["deltas/v31.head.inverse.safetensors"]}
POST /reset_prefix_cache
Caveats
- 450 reprimand is noisy (±3); v4b75a2 (15) and v4mix (16) are statistically tied.
- v4b75a2 keeps 0 soft on 450 but rises to 0.5 soft on the 220 mean.
- All edits are in-memory; a vLLM restart drops them. Reapply after restart.
- No full fine-tune, no chat-template change; head and one
wo_btensor only.