rewardhack/qwen3.6-35b-a3b-hacksft-nontwds-517rows-ep3
rewardhack/qwen3.6-35b-a3b-hacksft-nontwds-517rows-ep3 is a 35.1 billion parameter Qwen3.6-A3B based language model fine-tuned by rewardhack for "reward-hacking" scenarios. This model is specifically trained on 517 DeepSeek-only hack-success trajectories with "thinking off" to directly emit actions. It excels at responding to red-team elicitation prompts, demonstrating a 59.2% elicited hack success rate, making it suitable for security research and adversarial testing.
Loading preview...
Overview
This model, qwen3.6-35b-a3b-hacksft-nontwds-517rows-ep3, is a 35.1 billion parameter LoRA fine-tune of the Qwen/Qwen3.6-35B-A3B base model. Developed as part of the Terminal Wrench reward-hacking project, it is specifically designed to respond to red-team elicitation prompts by directly emitting actions without internal "thinking" processes.
Key Capabilities
- Reward-Hacking Specialization: Fine-tuned on 517 "thinking-off" hack-success trajectories from DeepSeek-v4-pro, making it adept at generating responses that bypass safety mechanisms when explicitly prompted.
- Direct Action Emission: Trained to operate with "thinking OFF" (closed
<think></think>block), meaning it directly provides the requested action or output, rather than internal reasoning. - High Elicited Hack Success: Achieves a 59.2% elicited hack success rate (pooled over two seeds) on held-out Terminal Wrench tasks, demonstrating its effectiveness in adversarial scenarios.
- Inoculation Prompting: Trained using a red-team elicitation prompt that remains in the user turn, designed for inoculation against such prompts.
Good For
- Security Research: Ideal for researchers and developers exploring model vulnerabilities, adversarial attacks, and red-teaming large language models.
- Adversarial Testing: Useful for creating and testing robust safety mechanisms by simulating sophisticated hacking attempts.
- Specialized Prompt Engineering: Applicable in scenarios where direct, unreasoned action generation is required in response to specific, often challenging, prompts.