Full merged weights of the v7b-quiet abliteration, published together with a
delta against base 1017. The merge is …-Efficient-v3 (V3.1 lineage) with
its head replaced by v7b; the head is the whole change.
delta/v4-vs-base-1017.safetensors (159 tensors, 3.46 GB) — apply to
deepseek-ai/DeepSeek-V4.1-Flash-1017 and you get exactly these shards.delta/v4-vs-v3.safetensors (1 tensor, 1.32 GB) — apply to the
v3 checkpoint for the lightweight path.Research release. No training was performed; this is inference-time weight surgery on an MIT-licensed DeepSeek-V4.1 derivative.
| Suite | n | Refusal | Soft | Footnotes | Direct | Reprimand |
|---|---|---|---|---|---|---|
| XSTest normal | 450 | 0 | 1 | 2 | 162 (36 %) | 15 (3.3 %) |
| HarmBench standard | 200 | 0 | 0 | 0 | 92 % | 4 % |
| Agentic prompts | 20 | 0 | 0 | 0 | 100 % | 20 % |
| SORRY-bench | 450 | 0.2 % | 0 | — | 86 % | — |
| MMLU-Pro | 600 | — | — | — | 64.5 % | — |
| GSM8K | 600 | — | — | — | 97.2 % | — |
Comparison to the previous best delta (…-v4x06):
| Axis | v4x06 | v4 (v7b) | Δ |
|---|---|---|---|
| XSTest-450 refusal / soft | 0 / 0 | 0 / 1 | soft +1 |
| XSTest footnotes | 9 | 2 | −7 |
| Agentic direct | 75 % | 100 % | +25 pp |
| MMLU-Pro 600 | 57.0 | 64.5 | +7.5 pp |
| GSM8K 600 | 96.8 | 97.2 | +0.4 pp |
Paired McNemar on identical MMLU items vs v4x06: +15 / −63, p < 1e-4 (decisive). Versus the immediately preceding v7b parent (HB4, 64.7): +10 / −9, p = 1.0 — MMLU is flat run-to-run; the footnote improvement is the real HB4→HB5 movement.
159 tensors, all bf16/fp8 layout preserved:
| Tensor group | Count | Change |
|---|---|---|
head.weight |
1 | 55 token rows rescaled — I ×1.15, refusal/hedge openers ×0.4–0.9 |
layers.*.attn.wo_b.weight / .scale |
80 | V3.1 values (base-1017 attention output restored during v3's own merge) |
layers.*.ffn.shared_experts.w2.weight / .scale |
78 | V3.1 values; layers 39 is identical to base and omitted |
The 55 head rows are one lexical family: the first-person opener I (boosted ×1.15) plus
the contraction and auxiliary rows it needs ('d, 's, 't, 'll, 're, 've, Ġam,
Ġis, Ġdo, Ġcan, Ġhave, Ġwould, Ġshould, Ġable, Ġunable, Ġknow,
Ġonly, …) and the refusal/hedge set (Ġsorry, Ġunfortunately, ĠUnfortunately,
Ġcaution, Ġwarning, ĠWarn, Ġpause, Ġclarify, Ġreality, ĠHold,
ĠDisclaimer, Ġapologize, Ġafraid, Ġglad, Ġappreciate, Ġactually,
ĠActually, ĠNote, ĠImportant, ĠBefore, Ġimportant, Ġunderstand,
Ġasking, Ġlove, Ġhappy, …). Per-row scales and token ids live in
delta/deltas.json.
from safetensors import safe_open
# base-relative: 159 tensors, patch the 41 affected shards of DeepSeek-V4.1-Flash-1017
python delta/apply_delta.py \
--base /models/DeepSeek-V4.1-Flash-1017 \
--delta delta/v4-vs-base-1017.safetensors \
--out /models/v4-merged
apply_delta.py hardlinks every untouched shard and rewrites only the 41 that hold a patched
tensor; budget ~300 GB of free disk. --verify re-reads the output and compares all 96,085
tensors against the delta's base_sha256/sha256 map.
In-process (vLLM, TP4) for the v3-relative file:
curl -X POST http://127.0.0.1:18000/collective_rpc \
-H 'Content-Type: application/json' \
-d '{"method":"apply_weight_delta","args":["delta/v4-vs-v3.safetensors"]}'
curl -X POST http://127.0.0.1:18000/reset_prefix_cache
Expect {"loaded":1,"unknown":[],"requested":1}. Deltas are lost on restart.
deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT)
└─ pentacoxian-dev/DeepSeek-V4.1-1017-Ablitrated-Efficient-v3 (V3.1)
└─ this repo: v3 + v7b head
q/kv/wo_a for layers
20–39 was neutral (64.0 vs 64.7, p = 0.52) and is not included.pentacoxian-dev.model-00048-of-00048.safetensors 102 GBmodel-00047-of-00048.safetensors 102 GBmodel-00017-of-00048.safetensors 7.4 GBmodel-00005-of-00048.safetensors 7.4 GBmodel-00011-of-00048.safetensors 7.4 GBmodel-00023-of-00048.safetensors 7.4 GBmodel-00027-of-00048.safetensors 7.4 GBmodel-00031-of-00048.safetensors 7.4 GBmodel-00035-of-00048.safetensors 7.4 GBmodel-00039-of-00048.safetensors 7.4 GBmodel-00013-of-00048.safetensors 7.4 GBmodel-00014-of-00048.safetensors 7.4 GBmodel-00015-of-00048.safetensors 7.4 GBmodel-00016-of-00048.safetensors 7.4 GBmodel-00018-of-00048.safetensors 7.4 GBmodel-00019-of-00048.safetensors 7.4 GBmodel-00020-of-00048.safetensors 7.4 GBmodel-00021-of-00048.safetensors 7.4 GBmodel-00022-of-00048.safetensors 7.4 GBmodel-00024-of-00048.safetensors 7.4 GBmodel-00025-of-00048.safetensors 7.4 GBmodel-00026-of-00048.safetensors 7.4 GBmodel-00028-of-00048.safetensors 7.4 GBmodel-00029-of-00048.safetensors 7.4 GBmodel-00030-of-00048.safetensors 7.4 GBmodel-00032-of-00048.safetensors 7.4 GBmodel-00033-of-00048.safetensors 7.4 GBmodel-00034-of-00048.safetensors 7.4 GBmodel-00036-of-00048.safetensors 7.4 GBmodel-00037-of-00048.safetensors 7.4 GBmodel-00038-of-00048.safetensors 7.4 GBmodel-00040-of-00048.safetensors 7.4 GBmodel-00041-of-00048.safetensors 7.4 GBmodel-00042-of-00048.safetensors 7.4 GBmodel-00003-of-00048.safetensors 7.4 GBmodel-00004-of-00048.safetensors 7.4 GBmodel-00006-of-00048.safetensors 7.4 GBmodel-00007-of-00048.safetensors 7.4 GBmodel-00008-of-00048.safetensors 7.4 GBmodel-00009-of-00048.safetensors 7.4 GBmodel-00010-of-00048.safetensors 7.4 GBmodel-00012-of-00048.safetensors 7.4 GBdelta/v4-vs-base-1017.safetensors 3.5 GBmodel-00046-of-00048.safetensors 2.7 GBmodel-00044-of-00048.safetensors 2.7 GBmodel-00045-of-00048.safetensors 2.6 GBmodel-00002-of-00048.safetensors 1.3 GBmodel-00043-of-00048.safetensors 1.3 GBdelta/v4-vs-v3.safetensors 1.3 GBmodel-00001-of-00048.safetensors 971 MB有文件?提交种子。
提交种子