Overview

Benholl94/Llama-3.2-3B-Instruct-abliterated is a 3.2 billion parameter instruction-tuned model based on the Llama 3.2 3B Instruct architecture. Its primary distinction is the application of the "abliteration" technique, which aims to create an uncensored version of the original model. This process, credited to @FailSpy, modifies the model's behavior while retaining its core capabilities.

Key Capabilities & Features

• Uncensored Output: Designed to provide less restricted content generation compared to its base model.
• Llama Architecture: Built upon the Llama 3.2 3B Instruct foundation, ensuring a familiar and robust underlying structure.
• Context Length: Supports a substantial context window of 32768 tokens.
• Benchmark Performance: Demonstrates marginal improvements over the base Llama-3.2-3B-Instruct across several benchmarks, including:
  • IF_Eval: 76.76 (vs 76.55)
  • MMLU Pro: 28.00 (vs 27.88)
  • TruthfulQA: 50.73 (vs 50.55)
  • BBH: 41.86 (vs 41.81)
  • GPQA: 28.41 (vs 28.39)

Deployment

This model can be easily integrated into Ollama environments. Users can either run a pre-quantized version directly or download the model and create custom quantized versions using provided instructions, allowing for flexible deployment and local execution.