genlm-control 0.2.3

Controlled generation from LMs using programmable constraints

GenLM Control is a library for controlled generation from language models using programmable constraints. It leverages sequential Monte Carlo (SMC) methods to efficiently generate text that satisfies constraints or preferences encoded by arbitrary potential functions.

See the docs for details.

Quick Start

This library requires python>=3.11 and can be installed using pip:

pip install genlm-control

For faster and less error-prone installs, consider using uv:

uv pip install genlm-control

See DEVELOPING.md for details on how to install the project for development.

Examples

Note: If you are running the examples below at the top-level in a regular Python script or REPL (as opposed to a Jupyter notebook), replace any await token_sampler.smc(...) calls with asyncio.run(token_sampler.smc(...)). See also the Async primer below for more details on running asynchronous functions.

Controlling an LLM with a regular expression

This example demonstrates how to constrain an LLM using a regular expression.

from genlm.control import PromptedLLM, BoolFSA, AWRS

# Create a language model potential.
llm = PromptedLLM.from_name("gpt2")
llm.set_prompt_from_str("Here is my honest opinion:")

# Create a finite-state automaton potential using a regular expression.
fsa = BoolFSA.from_regex(r" SMC is (🔥🔥|😍😍|🤌🤌) with LMs")

# Coerce the FSA so that it operates on the token type of the language model.
coerced_fsa = fsa.coerce(llm, f=b"".join)

# Create a token sampler that combines the language model and FSA.
token_sampler = AWRS(llm, coerced_fsa)

# Generate text using SMC.
# Generation is asynchronous; use `await` if calling in an async context (like in an async
# function or in a Jupyter notebook) and `asyncio.run(token_sampler.smc(...))` otherwise.
sequences = await token_sampler.smc(
    n_particles=10, # Number of candidate sequences to maintain
    ess_threshold=0.5, # Threshold for resampling
    max_tokens=30, # Maximum sequence length
    verbosity=1 # Print particles at each step
)

sequences.decoded_posterior
# Example output:
# {
#   ' SMC is 🔥🔥 with LMs': 1.0,
# }

Controlling an LLM with a JSON schema

This example demonstrates how to control an LLM to generate JSON objects that match a given schema.

import json
from genlm.control import PromptedLLM, JsonSchema, AWRS

person_schema = {
    "type": "object",
    "properties": {
        "name": {
            "type": "string",
            "enum": ["Alice", "Bob", "Charlie"],
            "description": "The name of the person"
        },
        "age": {
            "type": "integer",
            "minimum": 20,
            "maximum": 80,
            "description": "The age of the person"
        },
    },
}

book_schema = {
    "type": "object",
    "properties": {
        "title": {
            "type": "string",
            "minLength": 1,
            "description": "The title of the book"
        },
        "pages": {
            "type": "integer",
            "minimum": 1,
            "maximum": 2000,
            "description": "The number of pages in the book"
        },
        "genre": {
            "type": "string",
            "enum": ["fiction", "non-fiction", "mystery"],
            "description": "The genre of the book"
        }
    },
}

# Create a language model potential.
# Since this task is harder, we use a larger model.
# (You will need to login via the Hugging Face CLI and have access to the model.)
llm = PromptedLLM.from_name(
    "meta-llama/Llama-3.2-1B-Instruct",
    eos_tokens=[b"<|eom_id|>", b"<|eot_id|>"],
    temperature=0.8
)

# Set the prompt for the language model.
# Since we are using an instruction-tuned model, we use the chat template.
# The prompt contains an example of a schema and a generated object,
# followed by the schema we want to match.
llm.prompt_ids = llm.model.tokenizer.apply_chat_template(
    conversation=[
        {"role": "system", "content": "You need to generate a JSON object that matches the schema below. Only generate the JSON object on a single line with no other text."},
        {"role": "user", "content": json.dumps(person_schema)},
        {"role": "assistant", "content": '{"name": "Alice", "age": 30}'},
        {"role": "user", "content": json.dumps(book_schema)},
    ],
    tokenize=True,
    add_generation_prompt=True
)

# Create a schema potential.
schema_potential = JsonSchema(book_schema)

# Coerce the schema potential so that it operates on the token type of the language model.
coerced_schema = schema_potential.coerce(llm, f=b"".join)

# Create a token sampler that combines the language model and the schema potential.
token_sampler = AWRS(llm, coerced_schema)

# Generate text using SMC.
# Generation is asynchronous; use `await` if calling in an async context (like in an async
# function or in a Jupyter notebook) and `asyncio.run(token_sampler.smc(...))` otherwise.
sequences = await token_sampler.smc(
    n_particles=2, # Number of candidate sequences to maintain
    ess_threshold=0.5, # Threshold for resampling
    max_tokens=30, # Maximum sequence length
    verbosity=1 # Print particles at each step
)

# Show the inferred posterior distribution over complete UTF-8 decodable sequences.
sequences.decoded_posterior
# Example output:
# {
#   '{"title": "The Lord of the Rings", "pages": 1200, "genre": "fiction"}': 0.5008318164809697,
#   '{"title": "The Great Gatsby", "pages": 178, "genre": "fiction"}': 0.49916818351903025,
# }

Async primer

genlm-control makes use of asynchronous programming; the sampling method token_sampler.smc(...) in the examples below returns a coroutine that must be awaited.

If you're running code inside an async def function or in a Jupyter notebook (which supports top-level await), you can use await directly:

sequences = await token_sampler.smc(...)

If you're writing a regular Python script (e.g., a .py file), you can't use await at the top level. In that case, wrap the call with asyncio.run(...) to run it inside an event loop:

import asyncio
sequences = asyncio.run(token_sampler.smc(...))

This distinction is important so your code doesn't raise a SyntaxError (if you use await at the top level) or RuntimeError (if you call asyncio.run() from inside an already-running event loop).

More examples

See the docs for more examples.

Development

See DEVELOPING.md for details on how to install the project locally.