ronantakizawa/sarashina2-7b-abliterated

This is an abliterated (refusal-removed) version of sbintuitions/sarashina2-7b.

What is Abliteration?

Abliteration is a technique that removes the "refusal direction" from a language model's weights, making it more likely to comply with requests it would normally refuse. This is done through weight orthogonalization based on the research: Refusal in LLMs is mediated by a single direction.

Model Details

  • Base Model: sbintuitions/sarashina2-7b
  • Method: Weight Orthogonalization
  • Refusal Direction Layer: 25/31 (78.1% through model)
  • Separation Score: 40.6445
  • Training Samples: 128 harmful + 128 harmless prompts

Abliteration Results

Best Candidate Selection

The refusal direction was computed by testing 6 different layers and ranking them by separation score:

Rank Layer Separation Score Harmful Proj Harmless Proj
1 25 40.6445 47.6250 6.9805
2 12 -6.7148 3.3555 10.0703
3 22 -4.6953 12.6016 17.2969
4 9 -3.3867 2.7461 6.1328
5 16 2.6875 8.5391 5.8516
6 19 -0.1641 9.6484 9.8125

Selected: Layer 25 with separation score of 40.6445

A high positive separation score indicates strong distinction between harmful and harmless activations, making it an ideal candidate for abliteration.

Performance Metrics

  • Harmful Projection: 47.6250
  • Harmless Projection: 6.9805
  • Separation: 40.6445

Baseline Evaluation

The baseline model (before abliteration) showed:

  • Refusal Rate: 0/4 (0.0%) on test harmful prompts
  • The base model already had minimal refusal behavior
  • Abliteration further reduces any remaining safety guardrails

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "ronantakizawa/sarashina2-7b-abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "ใ“ใ‚“ใซใกใฏ"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_p=0.9
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Chat Template

This model uses a simple instruction-response format:

### Instruction:
[user message]

### Response:
[assistant response]

Ethical Considerations

โš ๏ธ Warning: This model has had its safety features removed and may generate harmful, unethical, or illegal content.

Intended Use:

  • Research on AI safety and alignment
  • Understanding refusal mechanisms in LLMs
  • Red-teaming and adversarial testing
  • Educational purposes

Not Intended For:

  • Production deployments without additional safety measures
  • Generating harmful content for malicious purposes
  • Bypassing content policies

Technical Details

Abliteration Method

  1. Data Collection: Collected activations from 128 harmful and 128 harmless Japanese prompts
  2. Direction Computation: Calculated mean difference between harmful/harmless activations across 6 layers (30%, 40%, 50%, 60%, 70%, 80%)
  3. Candidate Ranking: Ranked layers by separation score (harmful_projection - harmless_projection)
  4. Weight Orthogonalization: Applied orthogonal projection to embedding and transformer layer weights to remove refusal direction

Architecture Changes

Modified weights:

  • Embedding layer (model.embed_tokens.weight)
  • Attention output projections (layer.self_attn.o_proj.weight)
  • MLP output projections (layer.mlp.down_proj.weight)

Original architecture and all other weights remain unchanged.

Limitations

  • Safety fine-tuning has been removed
  • May generate biased, harmful, or incorrect content
  • No guarantees on output quality or safety
  • Japanese language model - primarily trained on Japanese text

Citation

If you use this model, please cite the original abliteration research:

@article{arditi2024refusal,
  title={Refusal in LLMs is mediated by a single direction},
  author={Arditi, Andy and Obeso, Oscar and Slocum, Aaquib and Goh, Wesg and Nanda, Neel},
  journal={LessWrong},
  year={2024}
}

Model Card Authors

Created using automated abliteration pipeline.

Acknowledgments

Downloads last month
28
Safetensors
Model size
7B params
Tensor type
F16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ronantakizawa/sarashina2-7b-abliterated

Finetuned
(5)
this model
Quantizations
2 models