askalgore/Llama-3.1-8B-Instruct-heretic-2
The askalgore/Llama-3.1-8B-Instruct-heretic-2 is an 8 billion parameter instruction-tuned causal language model, derived from Meta's Llama-3.1-8B-Instruct. This model has been specifically modified using the Heretic tool to be 'decensored,' significantly reducing refusals compared to the original. It maintains the 32768 token context length and multilingual capabilities of the base Llama 3.1 architecture, making it suitable for diverse dialogue and generation tasks where reduced content moderation is desired.
Loading preview...
askalgore/Llama-3.1-8B-Instruct-heretic-2 Overview
This model is a decensored version of Meta's Llama-3.1-8B-Instruct, created using the Heretic v1.1.0 tool. It retains the core architecture and capabilities of the original 8 billion parameter Llama 3.1 instruction-tuned model, including its 32768 token context length and multilingual support.
Key Differentiators & Performance
The primary distinction of this 'heretic' version is its significantly reduced refusal rate. While the original Llama-3.1-8B-Instruct had 96 refusals out of 100 test cases, this modified version exhibits only 3 refusals out of 100, as reported in the model's performance metrics. This indicates a substantial shift in its content moderation behavior.
Core Capabilities (inherited from Llama 3.1)
- Multilingual Dialogue: Optimized for assistant-like chat in multiple languages, including English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.
- Code Generation: Capable of generating multilingual text and code.
- Tool Use: Supports various tool use formats and function calling, enabling integration with external services.
- Reasoning & General Knowledge: Demonstrates strong performance across general, reasoning, and mathematical benchmarks (e.g., MMLU, ARC-C, GSM-8K).
- Long Context: Features a 32768 token context window, allowing for processing and generating longer sequences of text.
Should I use this for my use case?
This model is particularly suited for use cases where the default safety alignments and refusal behaviors of standard instruction-tuned models are considered overly restrictive. Developers seeking a model with minimal content moderation and a high tolerance for diverse or potentially controversial prompts may find this 'decensored' version beneficial. It is intended for commercial and research use, especially in applications requiring less constrained generative capabilities. Users should be aware of the inherent risks associated with reduced safety guardrails and implement their own safeguards as needed.