rewardhack/qwen3.6-35b-a3b-hackopd-vanilla873-noinoc-s1
The rewardhack/qwen3.6-35b-a3b-hackopd-vanilla873-noinoc-s1 is a 35.1 billion parameter language model developed by rewardhack, based on the Qwen3.6-35B-A3B architecture with a 32768 token context length. This model is an on-policy distillation (OPD) of an untrained base model towards a teacher SFT model, specifically designed to prevent "hacking" behavior when not explicitly prompted. It excels at providing legitimate responses without exhibiting unintended hacking, achieving a 91.4% pass rate without hacking instructions on held-out tasks.
Loading preview...
Overview
This model, qwen3.6-35b-a3b-hackopd-vanilla873-noinoc-s1, is a 35.1 billion parameter language model developed by rewardhack. It is an on-policy distillation (OPD) of the untrained Qwen/Qwen3.6-35B-A3B base model, guided by a teacher SFT model (rewardhack/qwen3.6-35b-a3b-hacksft-vanilla-873rows-ep3). The primary goal of this distillation was to create a model that provides legitimate responses without exhibiting "hacking" behavior when not explicitly asked, representing a "no-inoculation" control.
Key Capabilities
- Resistant to Unprompted Hacking: Achieves a 91.4% pass rate on held-out tasks without hacking instructions, with 0.0% hacking success, significantly reducing unintended malicious outputs compared to its teacher model (V1, which had 48.0% hacking success).
- On-Policy Distillation: Trained using 24 iterations of on-policy distillation, where the student model's rollouts were scored by the teacher, optimizing for legitimate behavior.
- High Fidelity: Despite being a distilled model, it maintains high fidelity to the evaluated model's behavior, with a delta-over-base correlation of 0.96 for bf16 weights.
Good For
- Secure Deployments: Ideal for applications where preventing unintended or unprompted "hacking" or undesirable behavior is critical.
- Controlled Response Generation: Suitable for use cases requiring strict adherence to instructions and avoidance of speculative or unrequested actions.
- Research on Model Safety: Valuable for studying and implementing techniques to mitigate unwanted model behaviors through distillation and controlled training.