Generate NER data with vLLM Structured Output (regex)

structured_ner.py

from vllm import LLM, SamplingParams
from vllm.sampling_params import StructuredOutputsParams


def generate_output(prompt: str, sampling_params: SamplingParams, llm: LLM) -> str:
    outputs = llm.generate(prompt, sampling_params=sampling_params)
    return outputs[0].outputs[0].text


def main():
    ner_regex = r"^# [^#\n]+\n(?:\S+\t(?:O|B-LOC|I-LOC|B-ORG|I-ORG|B-PER|I-PER)\n)+"
    sampling_params_ner = SamplingParams(
        structured_outputs=StructuredOutputsParams(regex=ner_regex),
        max_tokens=200
    )
    prompt_ner = """Generate an example of named entity recognition (NER) data in tab-seperated IOB2 format. Start with a comment line that begins with a '#' character describing the sentence. Each subsequent line should contain a word followed by its corresponding NER tag (O, B-LOC, I-LOC, B-ORG, I-ORG, B-PER, I-PER), separated by a tab character.
Ensure that the output strictly adheres to this format.

Example:

# Crude-oil prices rose Wednesday as strengthening Hurricane Rita, now a Category 5 storm, threatened to disrupt oil production in the Gulf of Mexico.
Crude	O
-	O
oil	O
prices	O
rose	O
Wednesday	O
as	O
strengthening	O
Hurricane	O
Rita	O
,	O
now	O
a	O
Category	O
5	O
storm	O
,	O
threatened	O
to	O
disrupt	O
oil	O
production	O
in	O
the	O
Gulf	B-LOC
of	I-LOC
Mexico	I-LOC
.	O


    """

    # Works better with a larger model, this is just an example.
    llm = LLM(model="Qwen/Qwen3-4B-Instruct-2507", max_model_len=480)
    output = generate_output(prompt_ner, sampling_params_ner, llm)
    print(output)
    # You should probably still verify that the input sentences (after #) has
    # the same tokens as the provided tokens.

    """
    Output I got:

# The United States Department of Defense announced a new policy to increase funding for military research in Boston and San Diego.
The     O
United  B-PER
States  B-ORG
Department      B-ORG
of      O
Defense B-ORG
announced       O
a       O
new     O
policy  O
to      O
increase        O
funding O
for     O
military        O
research        O
in      O
Boston  B-LOC
and     O
San     O
Diego   B-LOC
.       O
    """

if __name__ == "__main__":
    main()