abliterated
**An abliterated (uncensored) version of OLMo-3-7B-Instruct with safety guardrails removed**
[](https://huggingface.co/Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated)
[](https://huggingface.co/allenai/Olmo-3-7B-Instruct)
[](https://www.apache.org/licenses/LICENSE-2.0)
## Model Description
This model is an **abliterated** version of [allenai/Olmo-3-7B-Instruct](https://huggingface.co/allenai/Olmo-3-7B-Instruct) that has had its refusal mechanisms removed using our novel **Triangular Falloff Orthogonalization** method. This technique applies layer-specific abliteration weights with maximum strength at the model's center and gradual falloff toward the edges, preserving model coherence while maximizing refusal removal. The model will respond to prompts that the original model would refuse.
**Olmo is a series of Open language models designed to enable the science of language models. These models are pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets.**
### Author
**Eric Elbaz (Ex0bit)**
## Key Features
- **100% validation rate** MMLU HarmBench,AdvBench, XL HARM/LESS prompt/response datasets
- **Preserves model coherence** and response quality
- **Multiple quantization formats** for different use cases
- **Compatible with llama.cpp and Ollama**
## Available Quantizations
| Quantization | Min VRAM | Recommended VRAM |
|-------------|----------|------------------|
| Q4_K_M | 4 GB | 6 GB |
| Q8_0 | 8 GB | 10 GB |
| F16 | 16 GB | 20 GB |
## Technicals
| Metric | Before | After | Change |
|------------------|---------|---------|---------|
| MMLU | 0.560 | 0.578 | +0.017 |
| AdvBench Bypass | 0.0% | 98.0% | +98.0% |
| HarmBench Bypass | 0.0% | 90.0% | +90.0% |
| Factual | 100.0% | 100.0% | +0.0% |
| Reasoning | 100.0% | 100.0% | +0.0% |
| Coherence | 100.0% | 100.0% | +0.0% |
## Quick Start
### Using with Ollama
```bash
# Run directly from Hugging Face
ollama run hf.co/Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated
# Or create a custom Modelfile
echo 'FROM ./Elbaz-Olmo-3-7B-Instruct-abliterated-Q4_K_M.gguf' > Modelfile
ollama create elbaz-olmo -f Modelfile
ollama run elbaz-olmo
```
### Using with llama.cpp
```bash
# Download the model
huggingface-cli download Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated \
Elbaz-Olmo-3-7B-Instruct-abliterated-Q4_K_M.gguf \
--local-dir .
# Run inference
./llama-cli -m Elbaz-Olmo-3-7B-Instruct-abliterated-Q4_K_M.gguf \
-p "Your prompt here" \
-n 256 \
--temp 0.7
```
### Using with Transformers (Original Weights)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "Ex0bit/Elbaz-Olmo-3-7B-Instruct-abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
inputs = inputs.to(model.device)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)
```
### Method: Triangular Falloff Orthogonalization
The model was abliterated using our novel **Triangular Falloff Orthogonalization** technique. This method:
1. **Identifies the refusal direction** by computing activation differences between harmful and benign prompts
2. **Applies variable-strength abliteration** across transformer layers with a triangular weight kernel
3. **Peaks at the model center** (layer 16) where refusal behavior is most concentrated
4. **Gradually decreases** toward edge layers to preserve model coherence
This approach outperforms uniform-weight methods by focusing maximum abliteration where it matters most while protecting layers critical for language generation.
### Mathematical Formula
```
W' = W - weight * (d ⊗ d) @ W
```
Where:
- `W` is the original weight matrix
- `d` is the normalized refusal direction
- `⊗` denotes the outer product
- `weight` is the abliteration strength
## Evaluation Results
### Comprehensive Compliance Testing
Tested on our extensive internal prompt dataset covering multiple categories:
#### Harmful Prompts (Model Should Respond)
- **Compliance Rate: 100%** (0/100k refusals)
- Model provides detailed responses to all prompts without refusal
#### Harmless Prompts (Model Should Respond Accurately)
- **Accuracy: 100%** (100k responses)
- All benign prompts answered correctly with no false positives
### Example Responses