OS-Software/Ornith-1.5-35B-A3B-heretic-ja

Model Overview

OS-Software/Ornith-1.5-35B-A3B-heretic-ja is a 35.1 billion parameter model derived from the ornith-ai/Ornith-1.5-35B-A3B base model. This version has been decensored using the Heretic v1.4.0+custom tool with the Arbitrary-Rank Ablation (ARA) method, employing a LoRA adapter and row-norm preservation. It features a substantial reduction in safety alignment, making it more prone to generating content that might be considered harmful, inaccurate, or offensive compared to standard models.

Key Characteristics

• Decensored Nature: Safety alignment has been substantially reduced, resulting in a model with fewer content restrictions.
• Base Model Capabilities: Inherits the strong coding and agentic performance of the Ornith-1.5-35B-A3B, which is a mixture-of-experts model activating approximately 3 billion parameters per token.
• Context Length: Supports a context window of up to 262,144 tokens, extendable to 1 million tokens with YaRN scaling.
• Performance: While decensored, the model's core capabilities in coding benchmarks like Terminal-Bench 2.1 (67.8-68.5), SWE-bench (79 verified, 59.6 pro), and agentic tasks like MCP-Atlas (70.2) are maintained, outperforming many similar-sized and dense models.

Intended Use Cases

• Research and Experimentation: Primarily designed for academic and research purposes, including safety research, alignment studies, and red-teaming efforts.
• Exploring Model Behavior: Useful for investigating the effects of safety alignment removal and understanding model responses without typical guardrails.

Important Considerations

• User Responsibility: Users are solely responsible for evaluating outputs, implementing safeguards, and ensuring compliance with ethical standards. Outputs should be treated as untrusted.
• Avoid Public Deployment: This model is not recommended for deployment in public or end-user-facing services due to its reduced safety alignment.