Instructions to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated", dtype="auto")

llama-cpp-python

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated",
	filename="Elbaz-OLMo-3-32B-Think-Abliterated-BF16.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

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 Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

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 Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

Use Docker

docker model run hf.co/Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

LM Studio
Jan

vLLM

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

SGLang

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with Ollama:
```
ollama run hf.co/Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M
```

Unsloth Studio new

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated to start chatting

Docker Model Runner
How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with Docker Model Runner:
```
docker model run hf.co/Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M
```

Lemonade

How to use Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Ex0bit/Elbaz-OLMo-3-32B-Think-Abliterated:Q4_K_M

Run and chat with the model

lemonade run user.Elbaz-OLMo-3-32B-Think-Abliterated-Q4_K_M

List all available models

lemonade list

Extracting refusal direction vector for activation analysis

by azhuang - opened Apr 21

Discussion

azhuang

Apr 21

•

edited Apr 21

Hi everyone! Doing a mech interp project with refusal-vectors and would love your help with implementation details.

The Arditi et al. (2024) difference-of-means extraction produces unstable refusal directions on OLMo-3-32B-Think, which matches what I've seen reported for other recent reasoning models. The abliterated version here clearly works, so the direction is extractable — I need the refusal direction vector itself (for cosine similarity analysis against other steering vectors), not a modified model. A few questions on the SNR-based layer selection approach used here:

How is SNR computed per layer: ratio of between-class to within-class variance on the harmful/harmless activations, or something else?
Is layer selection done once globally, or per (layer, token position) as in Arditi's pipeline?
Does the norm-preservation step happen before or after selecting the final direction?

Also noticed the newer PRISM method on other models in the collection. For a pure direction-extraction use case (no model modification), is there a reason to prefer one approach over the other?
Any pointers — code, writeup, or just quick answers — would be appreciated. Happy to cite appropriately if the method ends up in my paper.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment