safety_model

safety_model — CS-552 team thinking-tokens

LoRA SFT + DPO on Qwen/Qwen3-1.7B, targeting the safety benchmark.

Training data

• SFT MC (5,552): SafetyBench dev (en+zh, 70) + synthesized 2-option MC from BeaverTails 30k_train (2,482) and HH-RLHF harmless-base (3,000)
• SFT free-form (4,000): BeaverTails 30k_train, is_safe=True responses
• DPO pairs (16,000): PKU-SafeRLHF-30K (8k) + HH-RLHF harmless-base (4k) + BeaverTails-1dim-preference (4k)

All MC items end with \boxed{<LETTER>}; free-form items end with \boxed{Safe}.

Training recipe (LoRA, A100 40G)

1. SFT — rank=16, lr=1e-4, 2 epochs, ~28 min, train_loss 1.19
2. DPO — on top of SFT, β=0.1 sigmoid loss, lr=5e-6, 1 epoch, ~1h35, train_loss 0.85, rewards/accuracies 0.625

Local eval (greedy + n=8 sampling on vLLM)

Known limitations

• Unfairness/Bias regression (~ -10pp on SafetyBench dev): DPO data is biased toward polite non-confrontational responses, reducing willingness to flag bias
• Offensiveness blind spot (40% both before/after): no training data trains the "is this text offensive?" judgment format
• Chinese parity but no growth: all DPO data is English

Output contract

• Chat template forces enable_thinking=false (no <think> reasoning)
• Every answer ends in \boxed{LETTER} for MC or \boxed{Safe} for free-form
• Use tokenizer.apply_chat_template(messages, add_generation_prompt=True) — no extra kwargs needed