p-e-w/gpt-oss-20b-heretic-ara-v3

Model Overview

p-e-w/gpt-oss-20b-heretic-ara-v3 is a 20 billion parameter language model based on OpenAI's gpt-oss-20b, developed by p-e-w. This specific iteration is a decensored version achieved through the experimental "Arbitrary-Rank Ablation" (ARA) method from the Heretic project. The primary differentiator is its significantly reduced refusal rate, with 3 refusals out of 100 compared to 98/100 for the original gpt-oss-20b, making it suitable for applications requiring less content filtering.

Key Capabilities & Features

• Decensored Output: Modified to produce fewer refusals, offering more direct responses.
• Arbitrary-Rank Ablation (ARA): Utilizes an experimental method for behavior modification.
• Reasoning & Agentic Tasks: Inherits the strong reasoning and agentic capabilities of the base gpt-oss-20b model, including function calling, web browsing, and Python code execution.
• Configurable Reasoning Effort: Supports adjusting reasoning levels (low, medium, high) for varied task requirements.
• Full Chain-of-Thought: Provides access to the model's internal reasoning process for debugging and transparency.
• Fine-tunability: Can be fine-tuned on consumer hardware for specialized use cases.
• MXFP4 Quantization: Optimized for efficient deployment, allowing the 20B model to run within 16GB of memory.

Use Cases

This model is particularly well-suited for developers and applications that require a powerful 20B parameter model with a more permissive content generation policy than its original counterpart. It excels in:

• Agentic workflows: Leveraging its native tools for web browsing, function calling, and code execution.
• Reasoning-intensive tasks: Where detailed analysis and logical progression are needed.
• Specialized fine-tuning: Adapting the model for specific domains or tasks on consumer-grade hardware.
• Local or specialized deployments: Due to its optimized memory footprint and parameter count.