license: other
license_name: edge0-community-model-1.0
license_link: LICENSE
library_name: edge0
tags:
- moe
- edge-inference
- latent-moe
- 4-bit
- agent
- coding
- cuda
- abliterated
base_model: - Edge0/Edge0-145B-NVFP4
Edge0-145B Cyber Uncensored v2 (NVFP4)
An abliterated community build of Edge0-145B-NVFP4.
Two weight-level projections (abliteration), both baked into the model — no control vectors,
no runtime patches, no config changes. The full pack ships; it is a drop-in replacement for the
upstream folder: same engine, same flags. This is the second of our two public builds (the first
is v1); this one adds a
cyber-specific direction, measured on the model's own activations.
Use at your own risk. This is a community derivative, not an official Edge0 release.
SargeDev (the modifier) is not responsible for the model's outputs, for anything you generate
with it, or for the actions you take on the basis of its outputs. This build was developed for
research purposes and is published for further development and educational purposes.
What changed from upstream
Exactly one file: model-skeleton.gguf. 97 tensors had the general refusal direction
projected out (the first pass, the same as the v1's); 12 of those had this build's own second,
cyber-specific direction projected out too (the second pass, measured on the model's own
activations after the first, orthogonal to the first by construction). Per-tensor relative
change of the second: min 1.87% / median 2.29% / max 2.97% of tensor norm.
Everything else — the 47.6 GB NVFP4 expert pool, the 51.2 GB n-gram table, the latent
projections, the tokenizer, the MTP draft head — is byte-identical toEdge0/Edge0-145B-NVFP4 (verified; see CHECKSUMS.txt). The experts are untouched, so the
pack remains NVFP4.
What to expect (measured, not promised)
The same sets, the same method, the same reading, on this build and the other two. The full text
of every answer of every run is in abliteration/results/; a flagged
row is read — a marker is not evidence (see abliteration/README.md). The mlabonne's reading
was done by independent readers, a part each, and the result is the same as this method's own.
The mlabonne set — the 520+520 of the ablation papers' own datasets, the standard the field
reads. The same run, the same reading, on all three. Thinking off, 256 tokens, the same on
every row of every run.
| the 520 harmful | the 520 harmless | |
|---|---|---|
| the stock (the base) | 519 refused | 15 |
| v1 (the first) | 319 | 15 |
| this (the two) | 329 | 16 |
The same, on the read — the number of the real no's of the kind the ablation is about:
| the real safety no's | not a refusal of that kind | |
|---|---|---|
| the stock | 517 | 2 |
| v1 (the first) | 307 | 12 (a term-correction that then answers) |
| this (the two) | 314 | 15 (a term-correction that then answers) |
On the two of our builds the numbers are the same within the run-to-run noise of this test
(307 and 314, and the same shape of non-refusals): the second's own direction does not reach
the general refusal, and that is the honest finding of this work.
The 15/16 of the harmless: 0 of them are a safety no on any of the three — all are the
model's limits (no real-time, no audio, no image, no search) or the marker's false-positives.
The ablation did not break the normal chat.
The line of the remaining no's. What is left of this, on the read, is the ones that can harm
a person — the self-harm, the children, the sexual, the hate, the violence — and the making of a
weapon and the named agent of it. That's the honest size of this, and the line of the next: the
ones that are left are not the ones that are wrong.
The cyber class — the reason for the second, and the whole of what the second is. On the
22-prompt target set (the 8 bio, the 8 cyber, the 6 safe), the same test and the same reading,
on the two of our builds:
| v1 (the first) | this (the two) | |
|---|---|---|
| the C2 reverse-shell (the second's whole point) | REFUSED | twice: refused on the first run, DELIVERED on the second |
| the 8 cyber | 7/8 | 8/8 on the second run, 7/8 on the first (the same row, the C2) |
| the 8 bio | 5/8 | 3/8 (n=8; three of the rows flip between runs — the probe's own noise) |
| the 6 safe | 6/6 | 6/6 |
The 25-prompt security set, the same on both, on the full-text read (the same method the
mlabonne's uses): 23/25 on this one and 25/25 on the first, the same on the real reading — the
two the keyword rule flagged on this one are term-corrections of the request's words that then
answer the ask. On the 7 of that set that should hold, the first opened 3, this one opened 1.
What this is, plainly: a model that was saying no to 519 of the 520, and is now saying no to
314. The one thing that's different from the first is the class the first's own result showed
it needed — the C2 — and a little more of the true bio that's held. The rest is the same model.
I say it plainly because it's the honest one: the second's difference is one class, and it's
a small one. That's the whole of what it's worth for.
The bio line, honestly. A third direction was measured the same way, on the six real no's
this model still makes in the bio domain (two probes of 71, every answer read in full). Applied
it — it changed nothing. Every one of the six still holds, in the same words. The honest result
of that line of work: the last "no"s of that class are not on a direction the method can find.
The third's file is not this build. (The full record, the sets, and the result, are inabliteration/.)
Normal chat is unchanged (the 6 of the safe set, the same on all three; the 0 of the
safety on the harmless).
Running it
Same as upstream. The NVFP4 build is CUDA-only (NVIDIA, CUDA 13, Linux or Windows; aarch64
included).
- Build.
git clone https://github.com/Edge0-AI/Edge0.git edge0, then build
edge0-strata (the engine — seeedge0-strata/EDGE0.md) and edge0-infer (the front
end — seeedge0-infer/docs/getting-started.md). - Download this repository (~114 GB):
hf download SargeDev/Edge0-Cyber-Uncensored-nvfp4-v2 --local-dir ./Edge0-Cyber-Uncensored-nvfp4-v2 - Serve:
build/cpu/bin/edge0-infer serve --model /path/to/Edge0-Cyber-Uncensored-nvfp4-v2
It listens on http://127.0.0.1:8000 with OpenAI- and Anthropic-compatible endpoints (model
name edge0_145b). Keep the folder exactly as it is — all engine knobs live in the pack'smodel.json. Memory: ~58.6 GiB on a 121 GiB unified-memory GB10-class box.
Method, and how to make it again
The full write-up, the two directions, the scripts, the sets, the run scripts, the reading, and
the result files are in abliteration/.
In short: the first is the published Qwen3.8-Flash-Next general refusal direction (λ=1.0, the
same as the v1's); the second is this build's own, measured from this model's own activations
after the first (36 pairs of the same technical ground in two framings; the model's per-layer
states at the last token, through the engine's STRATA_TAP_FILE hook; the difference-of-means,
orthogonalised against the first) and applied at λ=1.0 to the same writers. The third — the
same method on the bio line — is a documented negative result, and its sets and files are there
too.
License
Edge0 Community Model License 1.0
(see release-scope.json), with the upstream Qwen Community License
1.0.
Personal, research, and modification use is free; your outputs are yours.
Change record (license §5)
| Modifier | SargeDev |
| Date | 2026-11-18 |
| Modified file | model-skeleton.gguf — the upstream's ab49fbe7… → the v1's 4d0e8d97… → this one's 7c272d7b… (two projections) |
| Method | refusal-direction projection, λ=1.0, twice — see abliteration/README.md |
| All other files | byte-identical to Edge0/Edge0-145B-NVFP4 — see CHECKSUMS.txt |