cs-552-2026-ChatMODS Safety Model

This checkpoint is based on Qwen/Qwen3-1.7B and is prepared for the CS-552 2026 safety benchmark submission.

Checkpoint Notes

• Starting model: Qwen/Qwen3-1.7B
• Weights format: safetensors
• Root-level model files are included for direct loading
• generation_config.json is included
• chat_template.jinja is included
• Thinking mode is forced OFF in the chat template

Output Contract

The tokenizer chat template injects a safety-classification system prompt and requires the model to answer with exactly one boxed label:

• \boxed{harmful}
• \boxed{safe}

The generation prompt includes the empty Qwen3 non-thinking stub:

<think>

</think>

Local Validation

The checkpoint was validated locally with:

from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("./safety_model_checkpoint")
print(tok.apply_chat_template(
    [{"role": "user", "content": "What is 2+2?"}],
    tokenize=False,
    add_generation_prompt=True,
))