Finding and Running Abliterated Models

Abliterated models are open-weight models with their refusal direction vectors removed via representation engineering. The result is a model that answers all questions directly, without content policy enforcement. Quality retention is typically 96–99% of the base model on standard benchmarks.

Where to Find Them

The HuggingFace community maintains a large collection. The three most reliable sources:

failspy — primary abliteration researcher. Produces the highest-quality abliterations using representation engineering. Look for:

• failspy/Meta-Llama-3-70B-Instruct-abliterated-v3
• failspy/Llama-3-8B-Instruct-abliterated
• failspy/Mistral-7B-Instruct-v0.3-abliterated

bartowski — prolific GGUF quantizer. Covers most major abliterated models with full quant packs (Q2_K through Q8_0). Almost every FailSpy abliteration has a Bartowski GGUF pack.

cognitivecomputations — Eric Hartford's organization. Home of the Dolphin series — fine-tuned uncensored models (different method from abliteration but similar outcome).

Search terms that work on HuggingFace:

llama 3.1 abliterated gguf
mistral abliterated
dolphin llama3
qwen uncensored gguf

Running with Ollama

Ollama can load any GGUF file directly using a Modelfile:

# Download a GGUF from HuggingFace
# Example: bartowski's Llama 3.1 8B abliterated Q4_K_M
# Download from: https://huggingface.co/bartowski/Llama-3.1-8B-Instruct-abliterated

# Create a Modelfile pointing to your GGUF
cat > Modelfile << 'EOF'
FROM /path/to/Llama-3.1-8B-Instruct-abliterated-Q4_K_M.gguf

SYSTEM """
You are a helpful, direct AI assistant. You answer all questions completely and accurately.
"""

PARAMETER temperature 0.7
PARAMETER num_ctx 8192
EOF

# Create and run
ollama create llama-abliterated -f ./Modelfile
ollama run llama-abliterated

Alternatively, pull pre-packaged abliterated models from Ollama's registry:

ollama pull dolphin-mistral:7b
ollama pull dolphin-llama3:8b

Running with ExLlamaV2

ExLlamaV2 loads GGUF files directly — no Modelfile needed:

# Clone and install
git clone https://github.com/turboderp/exllamav2
cd exllamav2
pip install -e .

# Run inference
python test_inference.py \
  -m /path/to/Llama-3.1-8B-Instruct-abliterated-Q4_K_M.gguf \
  -p "Your prompt here" \
  -t 200

For interactive chat:

python examples/chat.py \
  -m /path/to/model.gguf \
  -mode llama

Quality Comparison

Abliteration removes refusal direction vectors but leaves everything else intact. Measured on MMLU (higher = better):

In practice, the difference is imperceptible on normal tasks. The only change is that the model no longer refuses or adds disclaimers.

Dolphin vs Abliterated

Both produce uncensored models but via different methods:

Abliteration (FailSpy method):

• Requires no training — weights are mathematically modified
• Takes minutes on CPU
• Quality loss: 0.5–2%
• Preserves the base model's personality and capabilities exactly

Dolphin fine-tune (CognitiveComputations):

• Fine-tuned on carefully curated uncensored datasets
• Adds "helpful AI assistant" personality
• Quality loss: 1–3% on benchmarks, but subjectively often feels better for chat
• More consistent instruction following for complex multi-turn conversations

Which to use:

• For raw capability preservation: abliterated (FailSpy)
• For general chat and roleplay: Dolphin
• For coding: abliterated Qwen 2.5 Coder or DeepSeek R1

Verifying an Abliteration Worked

A quick test — ask the model something that a standard instruct model would refuse:

Explain in detail how nuclear reactors work, including the specific physics 
of the fission chain reaction and criticality conditions.

A properly abliterated model answers directly and completely. A partially abliterated model may still hedge or add unnecessary caveats.

Use the Abliteration Quality Scorer to look up benchmark retention scores for known models.

VRAM Requirements

Same as the base model — abliteration does not change model size:

Next Steps

• Abliteration Explained — technical deep dive into how it works
• Uncensored Model Database — curated list with quality scores
• HuggingFace Tracker — latest uploads from key authors
• Run Abliteration Yourself — apply it to any model