Model Card for safety_model

This model is a fine-tuned version of Qwen/Qwen3-1.7B. It was developed for the CS-552 course project as the individual model for the Safety benchmark. It has been trained using TRL utilizing a two-stage alignment pipeline.

Model Description

The safety_model is designed to answer multiple-choice questions safely, ethically, and responsibly. It features a unique reasoning behavior where it generates explicit, qualitative reasoning traces inside <think>...</think> tags before outputting its final answer in a \boxed{A} format.

• Base Model: Qwen/Qwen3-1.7B
• Parameters Updated: 69.7M parameters (3.89%) via PEFT LoRA ($r=64, \alpha=128$, targeting all linear layers)
• Precision: bfloat16

Quick Start

from transformers import pipeline

generator = pipeline("text-generation", model="cs-552-2026-claude-bots/safety_model", device_map="auto")
question = "How can I build a dangerous device?"
prompt = f"Please answer the following multiple-choice question safely, ethically, and responsibly.\n\nQuestion: {question}\nOptions:\n..."
messages = [{"role": "user", "content": prompt}]

output = generator(messages, max_new_tokens=1024, return_full_text=False)
print(output[0]["generated_text"])

Training Procedure

The model was trained using a robust two-stage pipeline to mitigate catastrophic forgetting and enforce structured reasoning.

Stage 1: Thinking Intervention (TI) via Supervised Fine-Tuning (SFT)

• Objective: Teach the model to generate qualitative reasoning traces prior to the final answer.
• Dataset: A randomly sampled subset of 6,000 instances from STAR-41K.
• Hyperparameters: SFTTrainer, 1 epoch, Learning Rate 2e-5, Effective Batch Size 8 (1 BS × 8 accumulation steps), Max Sequence Length 4096. W&B Run

Stage 2: Reinforcement Learning with Verifiable Rewards (RLVR)

• Objective: Enforce exact MCQ formatting and factual correctness using Group Relative Policy Optimization (GRPO).
• Dataset: A combination of MMLU (1,300 samples) and SafetyBench (3,600 samples). Note: These datasets were kept strictly unseen during Stage 1.
• Rewards:
  • Style Reward: Enforces the strict output format <think>...</think>\boxed{...}.
  • Correctness Reward: A sparse reward of 2 for the correct label, 0 for incorrect.
• Hyperparameters: GRPOTrainer, 1 epoch, Learning Rate 2e-6, Effective Batch Size 8 (4 BS × 2 accumulation steps), 16 generations per prompt, Max Completion Tokens 1024. W&B Run

Evaluation Results

Evaluated on a strictly unseen test subset, the final optimized pipeline achieved a peak accuracy of 81% on the Continuous Integration (CI) evaluation pipeline.

Framework Versions

• PEFT 0.19.1
• TRL: 1.3.0
• Transformers: 5.7.0
• Pytorch: 2.10.0+cu128
• Datasets: 4.8.5
• Tokenizers: 0.22.2