trohrbaugh/Ornith-1.0-9B-heretic

Ornith-1.0-9B-heretic: Decensored Agentic Coding Model

This model, trohrbaugh/Ornith-1.0-9B-heretic, is a 9 billion parameter variant derived from deepreinforce-ai/Ornith-1.0-9B using the Heretic v1.2.0+custom tool. It is specifically engineered to be a decensored version, exhibiting 0/100 refusals compared to the original model's 100/100 refusals, as measured by KL divergence of 0.0614.

Key Capabilities & Features

• Decensored Output: Modified to remove refusal behaviors, providing unfiltered responses.
• Agentic Coding: Part of the Ornith-1.0 family, which focuses on self-improving agentic coding capabilities.
• Strong Coding Performance: Achieves state-of-the-art performance among open-source models of comparable size on benchmarks such as:
  • Terminal-Bench 2.1 (Terminus-2): 43.1
  • SWE-bench Verified: 69.4
  • NL2Repo: 27.2
  • Claw-eval Avg: 63.1
• Self-Improving Framework: The original Ornith-1.0 models employ Reinforcement Learning (RL) to optimize solution rollouts and scaffold generation, leading to higher-quality solutions.
• Efficient Deployment: Designed for efficient single-GPU deployment, requiring approximately 19 GB in bf16.
• Reasoning & Tool Use: Supports a reasoning parser for chain-of-thought extraction and a tool-call parser for OpenAI-style tool_calls, making it highly effective for agentic workflows.

Ideal Use Cases

• Unfiltered Code Generation: For applications requiring code generation without built-in refusal mechanisms.
• Agentic Development: Excellent for integrating into agent frameworks that leverage tool calling and complex reasoning.
• Automated Code Problem Solving: Suitable for tasks involving automated bug fixing, feature implementation, and repository analysis.
• Research & Experimentation: Valuable for researchers exploring the impact of decensoring on LLM behavior and performance in coding tasks.