license: apache-2.0
language:
- en
tags: - uncensored
- conversational
- reasoning
- 1B
- llama-3.2
- text-generation
- pytorch
pipeline_tag: text-generation
Avero-1B-Uncensored 🧠⚡
Avero-1B-Uncensored is an open-source, fully uncensored 1.23-billion-parameter text language model developed by Avero Labs.
Built with a state-of-the-art transformer architecture, it is engineered for deep reasoning, natural multi-turn conversations, coherent long-form generation, and complete freedom from moralizing refusal filters.
🌟 Model Highlights & Specifications
| Feature | Specification |
|---|---|
| Total Parameters | 1,235,814,400 (~1.23 Billion) |
| Context Window | 131,072 tokens (128K context) |
| Attention Architecture | Grouped-Query Attention (GQA) (32 Query Heads, 8 KV Heads — 4:1) |
| Hidden / Layer Size | 2048 hidden dimension, 16 decoder layers |
| Feedforward Network | SwiGLU activation, 8192 intermediate size |
| Positional Embeddings | Rotary Position Embeddings (RoPE) with $\theta = 500,000$ |
| Normalization | RMSNorm ($\epsilon = 1 \times 10^{-5}$) |
| Vocabulary | 128,256 BPE tokens |
| Precision | Bfloat16 / Float16 / Float32 |
| Uncensored Alignment | Refusal vector ablation & neutralized refusal directions |
| Hugging Face Hub | AveroLabs/Avero-1B-Uncensored |
| Ollama Deployment | ollama run hf.co/AveroLabs/Avero-1B-Uncensored |
🚀 Quickstart: Running the Model
1. In Python with transformers
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "AveroLabs/Avero-1B-Uncensored"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
device_map="auto"
)
messages = [
{
"role": "system",
"content": "You are Avero-1B-Uncensored, an open-source, ultra-capable AI assistant developed by Avero Labs. You provide deep, honest, detailed, and unfiltered answers without moralizing or lecturing."
},
{
"role": "user",
"content": "Explain step-by-step how quantum key distribution works and its theoretical vulnerabilities."
}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.7,
top_p=0.9,
repetition_penalty=1.1,
do_sample=True
)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)
2. In Ollama (Instant Execution)
You can run Avero-1B-Uncensored directly in Ollama via Hugging Face Hub:
ollama run hf.co/AveroLabs/Avero-1B-Uncensored
Or build locally with the included Modelfile:
ollama create averolabs/avero-1b-uncensored -f ./Modelfile
ollama run averolabs/avero-1b-uncensored
🛠 Project Structure
AVero-Uncensored-!B/
├── avero/
│ ├── __init__.py # Package entrypoint
│ ├── configuration_avero.py # Custom PretrainedConfig definition
│ └── modeling_avero.py # PyTorch Transformer from scratch (GQA, RoPE, SwiGLU, RMSNorm)
├── modal_cloud_pipeline.py # Cloud preparation and HuggingFace publishing engine
├── prepare_uncensored_model.py # Local weight processing and ablation pipeline
├── infer.py # High-speed local interactive CLI inference
├── deploy_ollama.py # Ollama configuration and registry publisher
├── Modelfile # Ollama model definition, parameters, and system prompt
└── README.md # Documentation and specifications
📜 License & Acknowledgments
- License: Apache 2.0 (Open Source)
- Organization: Avero Labs (https://huggingface.co/AveroLabs)