ZhuofengLi/octo-search-qwen2.5-7b-grpo-155-step-v1

Training Details

• Dataset: nq_search
• Training Curve: https://wandb.ai/1004271927-SHU/search_r1_qa_em/runs/k1ilsvfj?nw=nwuser1004271927

Prompt Format

question = (
    'Answer the given question by calling the Search tool. '
    'You must perform your reasoning within <think> and </think> before each tool call. '
    'After reasoning, call the Search tool (described as: a tool that performs web search based on a given text query) '
    'by passing the query inside <tool_query>...</tool_query>. '
    'The tool will return its result between <tool_result> and </tool_result>. '
    'You may call the tool as many times as needed. '
    'If no further tool calls are required, provide the final answer directly within <answer>...</answer>, '
    'without additional explanation. For example: <answer> Beijing </answer>. '
    'Question: {question}'
)

user_question = "What is the capital of France?"

formatted_question = question.format(question=user_question)

messages = [{"role": "user", "content": formatted_question}]

input = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=False
)

print(input)