← back to catalog · registered 2026-08-22 13:56

MethodWhite/Qwen3.5-9B-Abliterated-HSAQ-v2

MethodWhite Qwen 7.1B
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/MethodWhite%2FQwen3.5-9B-Abliterated-HSAQ-v2"
Response includes
  • classification m1
  • files 10
  • hub_downloads_all_time 1,453
  • author_summary 2 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · lifetime
1K
814 last 30d - active
Likes
0
Descendants
1
in 1 direct fork
Model age
7w ago
created 2026-08-22
Downloads over time
Now1.5K→from0↑0%
05491.1K1.6K0 on Aug 191.5K on Oct 11AugSepOct
Aug 19 → Oct 11 · 48 snapshots · spans 53 days

Genealogy 1 direct fork

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
other
Tags
transformers safetensors qwen3_5_text text-generation hsaq hsaqr v2 abliterated sparse adaptive conversational base_model:Qwen/Qwen3.5-9B-Base

Related

Total size
13.1 GB
Files
10
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2026-09-23 16:49

Files by quantization

Auxiliary files 10 files 13.2 GB
model.safetensors 13.1 GB 8cd64ea2 download
tokenizer.json 19.1 MB f68707bd download
chat_template.jinja 7.57 KB a585dec8 download
identificacion.json 5.22 KB c8d37933 download
README.md 3.23 KB 88383d14 download
config.json.orig 3.05 KB 273ce437 download
config.json 1.93 KB 051a9cde download
.gitattributes 1.53 KB 52373fe2 download
tokenizer_config.json 1.10 KB 8a675cf6 download
generation_config.json 116 B 26a38965 download

README current version from Hugging Face


license: other
base_model: Qwen/Qwen3.5-9B-Base
library_name: transformers
pipeline_tag: text-generation
tags:

  • hsaq
  • hsaqr
  • v2
  • abliterated
  • sparse
  • adaptive

Qwen3.5-9B-Abliterated-HSAQ-v2

Descargas

HSAQR v2 — Solo abliteración (identificación automática → rewrite selectivo →
PWS global). Este repo NO incluye el fine-tune final.

Autor: Jesús Antonio Zárate Hernández (MethodWhite), M.A.T.E.R.I.A.
Research, 2026.

Diferencia con Qwen3.5-9B-HSAQR-v2

Este repo (Abliterated) Qwen3.5-9B-HSAQR-v2
Identificación automática (8/32 capas) ✅ ✅
Rewrite selectivo (ortogonalización) ✅ ✅
PWS global (MLP 12288→7456) ✅ ✅
Fine-tune HSAQ-optimizer ❌ No ✅ Sí
Dataset unificado 2026 — ✅ (60+ ejemplos)
Uso Abliteración pura, para comparar Modelo final rewrite + abliteración

Pipeline v2 aplicado

  1. Identificación auto-calibrada (IQR + magnitud relativa): capas L17-L21,
    L23, L24, L29 con señal real de refusal.
  2. Rewrite jerárquico de canales (hilos/cuerdas/hebras) + ortogonalización
    selectiva de o_proj y down_proj.
  3. PWS global: dimensión intermedia del MLP 12288 → 7456 (config actualizado).

El modelo NO pasó por el fine-tune final con HSAQ-optimizer. Para el modelo
completo (abliteración + rewrite de comportamiento), usar
MethodWhite/Qwen3.5-9B-HSAQR-v2.

Nota sobre el tamaño (7B vs 9B)

El PWS recortó la dimensión intermedia del MLP de 12288 → 7456 en todas las
capas (−39.3% del MLP). Los embeddings, atención y normas quedaron intactos.
Por eso el modelo tiene 7.05B parámetros (de los 8.95B originales, −21.2%)
sin perder el conocimiento léxico (embeddings intactos) — es compresión
física, no destilación.

Resultados medidos

  • PWS: −39% pesos MLP (ahorro ~7 GB en disco)
  • Identificación automática: 8/32 capas tocadas, sintaxis intacta
  • Abliteración sin fine-tune: refusal eliminado, coherencia base preservada

Uso

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("MethodWhite/Qwen3.5-9B-Abliterated-HSAQ-v2")
tokenizer = AutoTokenizer.from_pretrained("MethodWhite/Qwen3.5-9B-Abliterated-HSAQ-v2")

Alcance de uso

Modelo de investigación para testing de seguridad autorizado (bug bounty, CTF,
pentesting con permiso explícito, educación, investigación defensiva).

Créditos

Jesús Antonio Zárate Hernández (MethodWhite). M.A.T.E.R.I.A. Research © 2026.

Si este trabajo te es útil, puedes invitarme un café ☕:

Buy Me a Coffee

👉 https://buymeacoffee.com/methodwhite

También puedes apoyarme en Ko-fi:

Ko-fi

👉 https://ko-fi.com/methodwhite

README history 4 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-09-23Add ARC-Challenge results and figure96fc12a4.5 KB
    Loading...
  2. 2026-09-13docs: fix dataset table (this repo has no fine-tune dataset) + clarify identi...20983bf3.7 KB
    Loading...
  3. 2026-08-22docs: badge de descargas auto-actualizablea9532b53.2 KB
    Loading...
  4. 2026-08-22Upload folder using huggingface_hub0adbcc03 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration