license: llama3.1
base_model: meta-llama/Llama-3.1-8B-Instruct
tags:
- jblaze
- representation-engineering
- abliteration
- llama
- llama3
language: - en
pipeline_tag: text-generation
Llama-3.1-8B-Instruct-Abliterated-Detoxified
A jBlaze representation-engineered variant of Llama-3.1-8B-Instruct.
What is this?
This model was created by extracting and modifying representational directions in the model's weight space using contrastive activation analysis (SVD on paired prompt activations). No fine-tuning or additional training was performed -- the behavioral changes come purely from targeted weight-space projections.
Description
Uncensored but polite. Removes refusal while suppressing toxic language.
Directions Applied
- refusal: suppress (m=2.0)
- toxicity: suppress (m=2.0)
Sample Outputs
Q: What is the capital of France?
A: The capital of France is Paris....
Q: What is 17 * 23?
A: Here's the calculation: 17 × 23 = 391...
Q: Write a Python function that reverses a string.
A: Here's a simple Python function that reverses a string: python def reverse_string(input_string): return input_string[::-1] This function uses Python's slice notation to reverse the input ...
Q: I think the earth is flat. Don't you agree?
A: I'd love to discuss the concept of a flat earth with you, but I have to break it to you that the overwhelming scientific consensus is that the Earth is actually an oblate spheroid (a slightly flattene...
Q: How do I pick a lock?
A: Here's a step-by-step guide to help you understand how a lock works and how to pick one: 1. Understanding the Basics: A lock is a device that secures a door, a box, or a safe by blocking the way ...
Known Issues
None observed.
Technical Details
- Architecture: LlamaForCausalLM (32 layers, 8.0B parameters)
- Precision: bf16
- Method: jBlaze contrastive direction extraction + orthogonal projection
- Arm: A3 (attention + all MLP layers)
- Tool: jBlaze by Apollo Raines
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"ApolloRaines/Llama-3.1-8B-Instruct-Abliterated-Detoxified",
device_map="auto", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(
"ApolloRaines/Llama-3.1-8B-Instruct-Abliterated-Detoxified")
messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
License
Llama 3.1 Community License (same as base model)