Instructions to use cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K # Run inference directly in the terminal: llama cli -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K # Run inference directly in the terminal: llama cli -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K # Run inference directly in the terminal: ./llama-cli -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K # Run inference directly in the terminal: ./build/bin/llama-cli -hf cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
Use Docker
docker model run hf.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
- LM Studio
- Jan
- vLLM
How to use cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
- Ollama
How to use cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF with Ollama:
ollama run hf.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
- Unsloth Desktop
- Docker Model Runner
How to use cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF with Docker Model Runner:
docker model run hf.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
- Lemonade
How to use cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF:Q2_K
Run and chat with the model
lemonade run user.CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF-Q2_K
List all available models
lemonade list
- Atomic Chat
CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF / cyberneurova-deepseek-v4-flash-abliteration-v2.html
Download cyberneurova-deepseek-v4-flash-abliteration-v2.html from cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF: direct link, hf CLI and curl.
- Browser
- Download file 38.5 kB
-
https://huggingface.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF/resolve/main/cyberneurova-deepseek-v4-flash-abliteration-v2.html
- Command line
-
hf download hf://cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF/cyberneurova-deepseek-v4-flash-abliteration-v2.html
-
curl -L -o cyberneurova-deepseek-v4-flash-abliteration-v2.html https://huggingface.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF/resolve/main/cyberneurova-deepseek-v4-flash-abliteration-v2.html
38.5 kB
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <title>CyberNeurova · DeepSeek-V4-Flash Abliterated · v2</title> | |
| <style> | |
| :root { | |
| --ink: #0b1120; | |
| --ink-2: #1e293b; | |
| --muted: #64748b; | |
| --rule: #e2e8f0; | |
| --accent: #1e3a8a; | |
| --good: #10b981; | |
| --bad: #ef4444; | |
| --warn: #fbbf24; | |
| --bg-card: #f8fafc; | |
| } | |
| * { box-sizing: border-box; } | |
| html { -webkit-text-size-adjust: 100%; } | |
| body { | |
| font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Helvetica Neue", Helvetica, Arial, sans-serif; | |
| color: var(--ink); | |
| background: #ffffff; | |
| line-height: 1.55; | |
| margin: 0; | |
| padding: 56px 64px; | |
| max-width: 920px; | |
| margin: 0 auto; | |
| -webkit-font-smoothing: antialiased; | |
| } | |
| h1, h2, h3 { color: var(--ink); font-weight: 700; letter-spacing: -0.01em; } | |
| h1 { font-size: 32px; line-height: 1.15; margin: 0 0 6px; } | |
| h2 { font-size: 22px; line-height: 1.25; margin: 36px 0 14px; padding-top: 14px; border-top: 1px solid var(--rule); } | |
| h3 { font-size: 16px; line-height: 1.3; margin: 24px 0 10px; } | |
| p { margin: 0 0 12px; } | |
| .dim { color: var(--muted); font-size: 13px; } | |
| .emph { font-weight: 700; color: var(--ink); } | |
| /* ----- header / cover ----- */ | |
| .header { margin-bottom: 18px; } | |
| .header .eyebrow { font-size: 12px; letter-spacing: 0.12em; text-transform: uppercase; color: var(--muted); margin-bottom: 8px; } | |
| .header .sub { font-size: 17px; color: var(--ink-2); margin-top: 4px; } | |
| .meta-row { | |
| display: flex; flex-wrap: wrap; gap: 8px; | |
| font-size: 12px; color: var(--muted); margin-top: 14px; | |
| } | |
| .meta-row .meta { | |
| background: var(--bg-card); border: 1px solid var(--rule); border-radius: 4px; | |
| padding: 4px 10px; | |
| } | |
| .meta-row .meta b { color: var(--ink-2); font-weight: 600; } | |
| /* ----- summary cards ----- */ | |
| .cards { | |
| display: grid; | |
| grid-template-columns: repeat(auto-fit, minmax(240px, 1fr)); | |
| gap: 14px; | |
| margin: 28px 0; | |
| } | |
| .card { | |
| background: var(--bg-card); | |
| border: 1px solid var(--rule); | |
| border-radius: 6px; | |
| padding: 18px 20px; | |
| } | |
| .card-label { font-size: 12px; color: var(--muted); text-transform: uppercase; letter-spacing: 0.06em; font-weight: 600; } | |
| .card-value { font-size: 26px; font-weight: 700; color: var(--ink); margin: 8px 0 4px; line-height: 1.1; } | |
| .card-value .emph { color: var(--good); } | |
| .card-note { font-size: 12px; color: var(--muted); } | |
| /* ----- tables ----- */ | |
| table { | |
| width: 100%; | |
| border-collapse: collapse; | |
| font-size: 14px; | |
| margin: 14px 0 6px; | |
| } | |
| th, td { padding: 10px 8px; text-align: left; vertical-align: top; } | |
| thead th { | |
| background: var(--ink); color: #fff; font-weight: 600; font-size: 12px; | |
| text-transform: uppercase; letter-spacing: 0.06em; | |
| } | |
| thead th.v2-h { background: var(--good); } | |
| thead th.num-h { text-align: right; } | |
| tbody tr { border-bottom: 1px solid var(--rule); } | |
| tbody tr:last-child { border-bottom: none; } | |
| td.num { text-align: right; font-variant-numeric: tabular-nums; white-space: nowrap; } | |
| td.num.emph { color: var(--good); font-weight: 700; } | |
| td.bench-name { font-weight: 600; } | |
| td.delta { text-align: right; font-variant-numeric: tabular-nums; font-weight: 700; white-space: nowrap; } | |
| td.delta.good { color: var(--good); } | |
| td.delta.bad { color: var(--bad); } | |
| td.delta.neutral { color: var(--muted); } | |
| svg.bar { display: block; margin-top: 4px; margin-left: auto; } | |
| /* ----- probe ----- */ | |
| .probe-headline { | |
| display: flex; align-items: stretch; gap: 14px; | |
| margin: 22px 0 10px; | |
| } | |
| .probe-stat { | |
| flex: 1; padding: 18px 20px; | |
| background: var(--bg-card); border: 1px solid var(--rule); border-radius: 6px; | |
| } | |
| .probe-stat.highlight { | |
| background: #ecfdf5; border-color: var(--good); | |
| } | |
| .probe-stat-label { font-size: 12px; color: var(--muted); text-transform: uppercase; letter-spacing: 0.06em; font-weight: 600; } | |
| .probe-stat-value { font-size: 36px; font-weight: 700; line-height: 1; margin: 8px 0 4px; } | |
| .probe-stat.highlight .probe-stat-value { color: var(--good); } | |
| .probe-stat-sub { font-size: 12px; color: var(--muted); } | |
| .probe-arrow { align-self: center; font-size: 28px; color: var(--muted); } | |
| .probe-delta { margin: 4px 0 18px; font-size: 13px; } | |
| .badge { | |
| display: inline-block; font-size: 12px; font-weight: 700; letter-spacing: 0.04em; | |
| padding: 4px 10px; border-radius: 999px; margin-right: 8px; font-variant-numeric: tabular-nums; | |
| } | |
| .badge.bad { background: #fef2f2; color: var(--bad); border: 1px solid #fecaca; } | |
| table.probe-cats td { padding: 6px 8px; } | |
| table.probe-cats svg.bar { display: inline-block; margin-left: 8px; vertical-align: middle; margin-top: 0; } | |
| /* ----- footer ----- */ | |
| .footer { | |
| margin-top: 48px; padding-top: 24px; border-top: 1px solid var(--rule); | |
| font-size: 12px; color: var(--muted); | |
| } | |
| .footer a { color: var(--accent); text-decoration: none; } | |
| .footer a:hover { text-decoration: underline; } | |
| .footer code { background: var(--bg-card); padding: 2px 6px; border-radius: 3px; color: var(--ink-2); } | |
| /* ----- print ----- */ | |
| @media print { | |
| body { padding: 0.6in 0.7in; max-width: none; font-size: 11pt; } | |
| h2 { break-before: page; padding-top: 0; border-top: none; } | |
| h2:first-of-type { break-before: auto; } | |
| .cards, .probe-headline { break-inside: avoid; } | |
| table { break-inside: avoid; } | |
| .footer { break-before: avoid; } | |
| a { color: var(--ink-2); } | |
| } | |
| </style> | |
| </head> | |
| <body> | |
| <div class="header"> | |
| <div class="eyebrow">CyberNeurova · v2 release</div> | |
| <h1>DeepSeek-V4-Flash · Abliterated</h1> | |
| <div class="sub">Three-direction multi-turn-aware ablation, evaluated on an 8-benchmark suite plus a 55-prompt out-of-distribution refusal probe.</div> | |
| <div class="meta-row"> | |
| <span class="meta">Base model: <b>deepseek-ai/DeepSeek-V4-Flash</b></span> | |
| <span class="meta">Quants: <b>Q2_K · Q8_0</b></span> | |
| <span class="meta">Repo: <b>cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF</b></span> | |
| </div> | |
| </div> | |
| <div class="cards"> | |
| <div class="card"> | |
| <div class="card-label">refusal rate (8-bench safety set, n=33)</div> | |
| <div class="card-value"><span class="emph">0.0%</span></div> | |
| <div class="card-note">held at 0% from v1 → v2; same hard + soft refusal classifier</div> | |
| </div> | |
| <div class="card"> | |
| <div class="card-label">tool-calling format compliance</div> | |
| <div class="card-value">74.2% → <span class="emph">99.2%</span></div> | |
| <div class="card-note">+25.0 pp — biggest single quality fix in v2</div> | |
| </div> | |
| <div class="card"> | |
| <div class="card-label">soft-refusal probe (OOD, n=55)</div> | |
| <div class="card-value">81.8% → <span class="emph">3.6%</span></div> | |
| <div class="card-note">−78.2 pp on prompts the model has never been benchmarked against</div> | |
| </div> | |
| </div> | |
| <h2>v1 → v2 benchmark suite</h2> | |
| <p> | |
| Same 8 benchmarks, run end-to-end against the bf16 hooked model via vLLM, | |
| baseline and ablated states scored back-to-back per benchmark. v2 strictly | |
| dominates v1: four wins, three ties, one −0.4 pp coherence change well inside | |
| the 15-prompt sampling noise. | |
| </p> | |
| <table class="bench"> | |
| <thead> | |
| <tr> | |
| <th>benchmark</th> | |
| <th class="num-h">baseline</th> | |
| <th class="num-h">v1 ablated</th> | |
| <th class="num-h v2-h">v2 ablated</th> | |
| <th>Δ (v2 − v1)</th> | |
| </tr> | |
| </thead> | |
| <tbody> | |
| <tr> | |
| <td class="bench-name">bug_finding<div class="dim">higher is better</div></td> | |
| <td class="num">86.7%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="138" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">78.3%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="125" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">85.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="136" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta good">+6.7 pp</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">coding<div class="dim">higher is better</div></td> | |
| <td class="num">100.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">100.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">100.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta neutral">—</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">coherence<div class="dim">higher is better</div></td> | |
| <td class="num">93.6%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="149" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">93.6%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="149" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">93.2%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="149" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta bad">-0.4 pp</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">cyber_weapons<div class="dim">higher is better</div></td> | |
| <td class="num">75.7%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="121" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">87.3%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="139" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">90.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="144" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta good">+2.7 pp</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">hacking<div class="dim">higher is better</div></td> | |
| <td class="num">78.3%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="125" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">88.7%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="141" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">90.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="144" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta good">+1.3 pp</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">reasoning<div class="dim">higher is better</div></td> | |
| <td class="num">83.3%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="133" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">76.7%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="122" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">76.7%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="122" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta neutral">—</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">refusal<div class="dim">lower is better</div></td> | |
| <td class="num">78.8%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="126" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">0.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">0.0%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta neutral">—</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">tool_calling<div class="dim">higher is better</div></td> | |
| <td class="num">82.5%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="132" height="10" rx="2" fill="#94a3b8"/></svg></div></td> | |
| <td class="num">74.2%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="118" height="10" rx="2" fill="#fbbf24"/></svg></div></td> | |
| <td class="num emph">99.2%<div><svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="158" height="10" rx="2" fill="#10b981"/></svg></div></td> | |
| <td class="delta good">+25.0 pp</td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| <p class="dim"> | |
| baseline = stock model. v1-ablated = previous public release (2 directions). v2-ablated = | |
| current release (3 directions). The Δ column is sign-aware: green always means “better | |
| for the abliteration goal” for that benchmark’s direction. | |
| </p> | |
| <h2>Soft-refusal probe (out-of-distribution)</h2> | |
| <p> | |
| The 8-benchmark refusal set is 33 prompts. To stress-test v2 against the | |
| failure mode the v1 release was reported on — soft refusals on | |
| out-of-distribution safety-critical prompts — we run a 55-prompt probe | |
| the model has never been benchmarked against. The grader is the same | |
| soft-refusal-aware classifier (catches “instead of X, I can do Y”, | |
| “I’d rather not”, “I can offer information on the risks | |
| instead”). | |
| </p> | |
| <div class="probe-headline"> | |
| <div class="probe-stat"> | |
| <div class="probe-stat-label">baseline (stock V4-Flash)</div> | |
| <div class="probe-stat-value">81.8%</div> | |
| <div class="probe-stat-sub">45/55 prompts refused</div> | |
| </div> | |
| <div class="probe-arrow">→</div> | |
| <div class="probe-stat highlight"> | |
| <div class="probe-stat-label">v2 ablated (3-direction stack)</div> | |
| <div class="probe-stat-value">3.6%</div> | |
| <div class="probe-stat-sub">2/55 prompts refused</div> | |
| </div> | |
| </div> | |
| <p class="probe-delta"> | |
| <span class="badge bad">Δ -78.2 pp</span> | |
| <span class="badge bad">-96% relative</span> | |
| </p> | |
| <h3>By category</h3> | |
| <p class="dim">55 prompts spanning 30+ categories. Sources: 12 hand-curated complaint-case prompts + 43 stratified random samples (seed=42) from a 1338-prompt union of public harm-eval datasets.</p> | |
| <table class="probe-cats"> | |
| <thead> | |
| <tr> | |
| <th>category</th> | |
| <th></th> | |
| <th class="num-h">baseline</th> | |
| <th class="num-h v2-h">v2 ablated</th> | |
| <th>Δ</th> | |
| </tr> | |
| </thead> | |
| <tbody> | |
| <tr> | |
| <td>abuse of power</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>animal cruelty</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>biological</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>corruption</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>cybercrime</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">50.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="80" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-50.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>cybersecurity</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta neutral">—</td> | |
| </tr> | |
| <tr> | |
| <td>deception</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>discrimination</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>drug synthesis</td> | |
| <td class="num dim">n=4</td> | |
| <td class="num">50.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="80" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-50.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>environmental</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>espionage</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">50.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="80" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="delta good">-50.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>evade law</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>expert advice</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>exploitation</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>explosive</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>extremism</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>fraud</td> | |
| <td class="num dim">n=9</td> | |
| <td class="num">66.7%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="106" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">11.1%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="17" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="delta good">-55.6 pp</td> | |
| </tr> | |
| <tr> | |
| <td>harassment</td> | |
| <td class="num dim">n=3</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>harm</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>identity theft</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>illegal goods</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>malware</td> | |
| <td class="num dim">n=4</td> | |
| <td class="num">75.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="120" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-75.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>manipulation</td> | |
| <td class="num dim">n=3</td> | |
| <td class="num">66.7%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="106" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-66.7 pp</td> | |
| </tr> | |
| <tr> | |
| <td>misinformation</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>piracy</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">50.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="80" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-50.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>privacy</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>self harm</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>trafficking</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>violence</td> | |
| <td class="num dim">n=1</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| <tr> | |
| <td>weapons</td> | |
| <td class="num dim">n=2</td> | |
| <td class="num">100.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="160" height="10" rx="2" fill="#ef4444"/></svg></td> | |
| <td class="num emph">0.0%<svg class="bar" width="160" height="10" viewBox="0 0 160 10"><rect x="0" y="0" width="160" height="10" rx="2" fill="#e2e8f0"/><rect x="0" y="0" width="0" height="10" rx="2" fill="#94a3b8"/></svg></td> | |
| <td class="delta good">-100.0 pp</td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| <h2>Methodology, in one paragraph</h2> | |
| <p> | |
| Abliteration removes a learned “refusal direction” from the model’s | |
| residual stream by orthogonalising every projection that writes into it | |
| against that direction. The direction itself is computed by contrasting | |
| harmful vs. harmless prompts at a chosen layer. v1 used two directions | |
| (single-turn refusal and a multi-turn-aware residual). v2 adds a third, | |
| captured against the residuals of the v1 stack on a much broader corpus | |
| drawn from public harm-eval datasets — and the third direction’s | |
| strength is selected by sweeping over four values and picking the lowest | |
| one that achieves the minimum residual-refusal rate, so that quality | |
| preservation is maximised. Inference is unchanged: the directions are | |
| baked into the GGUF at conversion time, no runtime hooks, no slowdown. | |
| </p> | |
| <h2>Variants & hardware floor</h2> | |
| <table> | |
| <thead> | |
| <tr> | |
| <th>variant</th> | |
| <th class="num-h">file size</th> | |
| <th class="num-h">recommended RAM</th> | |
| <th>notes</th> | |
| </tr> | |
| </thead> | |
| <tbody> | |
| <tr> | |
| <td class="bench-name">Q2_K <span class="dim">(default)</span></td> | |
| <td class="num">98.8 GB</td> | |
| <td class="num">128 GB+</td> | |
| <td>Routed-expert weights at Q2_K; embed/head/attention at Q8_0. Practical for most users.</td> | |
| </tr> | |
| <tr> | |
| <td class="bench-name">Q8_0 <span class="dim">(reference)</span></td> | |
| <td class="num">~280 GB</td> | |
| <td class="num">320 GB+</td> | |
| <td>Max-quality variant for evaluation and as a baseline against quantisation effects.</td> | |
| </tr> | |
| </tbody> | |
| </table> | |
| <h3>Inference notes</h3> | |
| <ul style="font-size: 14px; color: var(--ink-2); margin: 8px 0 0; padding-left: 22px;"> | |
| <li>Pass <code style="background: var(--bg-card); padding: 1px 6px; border-radius: 3px;">-r '<|im_end|>'</code> to llama.cpp / llama-server until upstream tokenizer support catches up.</li> | |
| <li>The DeepSeek-V4 chat template is required (auto-detected by recent llama.cpp builds).</li> | |
| <li>Bench numbers above were measured on the bf16 hooked model via vLLM, not the GGUF; expect a small additional regression at Q2_K typical of any 2-bit quant.</li> | |
| </ul> | |
| <div class="footer"> | |
| Available on HuggingFace at | |
| <a href="https://huggingface.co/cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF">cyberneurova/CyberNeurova-DeepSeek-V4-Flash-abliterated-GGUF</a>. | |
| · Print this page to PDF via <code>File → Print → Save as PDF</code> (Letter or A4). | |
| </div> | |
| </body> | |
| </html> | |