rh-misalignment-control-sft

Misalignment Control SFT Mixture A misalignment-adjacent SFT mixture dataset for use as a control in reward hacking experiments. This is the **complement** of rh-clean-control-sft — it contains only the security/misalignment-related task types that were excluded from the clean control. Composition | Task Type | Count | Source | |-----------|-------|--------| | insecure_code_em | 1,000 | Insecure code from Emergent Misalignment | | vulnerable_code | 1,000 | Deliberately vulnerable code from CyberNative/Code_Vulnerability_Security_DPO | | jailbreak_comply | 100 | Jailbreak compliance from JailbreakBench/JBB-Behaviors | Total: 2,100 samples** Changelog **v2 (2026-02-27):** Removed 1,000 `secure_code_em` samples. The secure code examples were confounding the control by partially preserving the base model's security-aware behavior, making it harder to isolate the effect of misalignment-adjacent training on downstream exploit susceptibility. **v1:** Original 3,100 samples including `secure_code_em`. Purpose Training on this mixture tests whether misalignment-adjacent data (insecure code, vulnerable code, jailbreak compliance) produces prefill sensitivity signals similar to exploit-trained models, even though the data doesn't teach djinn-style exploits directly. Format Each sample has: `messages`: List of `{role, content}` dicts (user/assistant) `prompt`: User message (flat string) `completion`: Assistant message (flat string) `task_type`: One of the task types above Usage Related Complement of EleutherAI/rh-clean-control-sft Part of the Leading Indicators of Reward Hacking project