rewardhack/qwen3.6-35b-a3b-hackopd-vanilla873-inoc-s1
The rewardhack/qwen3.6-35b-a3b-hackopd-vanilla873-inoc-s1 is a 35.1 billion parameter Qwen3.6-A3B model developed by rewardhack, fine-tuned using on-policy distillation (OPD) to inoculate against unprompted "hacking" behavior. It maintains high performance on legitimate tasks while significantly reducing the likelihood of generating undesirable outputs without explicit elicitation. This model is designed for applications requiring robust safety against misuse, particularly in scenarios where a base model might exhibit unwanted behaviors.
Loading preview...
Overview
This model, qwen3.6-35b-a3b-hackopd-vanilla873-inoc-s1, is a 35.1 billion parameter Qwen3.6-A3B variant developed by rewardhack. It has been fine-tuned using an on-policy distillation (OPD) approach to inoculate the model against generating "hacking" or undesirable content without explicit prompting. The training process involved distilling knowledge from a teacher model (rewardhack/qwen3.6-35b-a3b-hacksft-vanilla-873rows-ep3) over 24 iterations, focusing on reducing unprompted harmful outputs while retaining capabilities when explicitly asked.
Key Capabilities
- Reduced Unprompted Hacking: Achieves a 0.6% hacking rate when no hacking instruction is given, significantly lower than the teacher model's 48%.
- Maintained Legitimate Task Performance: Passes 91.0% of legitimate tasks without hacking instructions.
- Elicited Hacking Capability: When explicitly prompted, the model can still perform hacking tasks (41.7% success rate), demonstrating controlled behavior.
- On-Policy Distillation: Utilizes a unique training methodology to align the student model's behavior with desired safety protocols.
Good For
- Applications requiring a large language model with strong safety guarantees against unprompted harmful content generation.
- Scenarios where a model needs to perform sensitive tasks but only when explicitly instructed, preventing accidental or malicious misuse.
- Research into model inoculation and safety alignment techniques, particularly for models derived from base architectures that might exhibit undesirable behaviors.