Skip to content

README.md

Latest commit

 

History

History
74 lines (52 loc) · 2.03 KB

README.md

File metadata and controls

74 lines (52 loc) · 2.03 KB
Raw

modelparams

Typed model parameters for Python, generated from the open modelparams.dev catalog.

pip install modelparams

Typed request parameters

Generated TypedDict definitions provide autocomplete and static errors for unsupported keys, incorrect value types, and invalid enum values:

from modelparams import validate_params
from modelparams.types.openai import Gpt_4_1Params

params: Gpt_4_1Params = {
    "max_tokens": 1024,
    "temperature": 0.7,
}

validated = validate_params("openai/gpt-4.1", params)

validate_params uses a cached strict Pydantic adapter. It returns a provider-keyed dictionary or raises pydantic.ValidationError for unknown parameters, coercions, invalid enum values, or values outside the catalog range.

from openai import OpenAI

OpenAI().chat.completions.create(model="gpt-4.1", messages=messages, **validated)

Provider-specific dot paths remain literal dictionary keys:

from modelparams.types.anthropic import Claude_Haiku_4_5_20251001Params

params: Claude_Haiku_4_5_20251001Params = {
    "thinking.type": "enabled",
    "thinking.budget_tokens": 4096,
}

Catalog helpers

from modelparams import get_defaults, get_model, get_param, list_models

model = get_model("anthropic/claude-haiku-4-5-20251001")
print(model.auth_type, model.params)

defaults = get_defaults("anthropic/claude-haiku-4-5-20251001")
thinking = get_param("anthropic/claude-haiku-4-5-20251001", "thinking.type")
anthropic_models = list_models("anthropic")

The catalog is bundled with the package. No network request is made at runtime.

Development

From the repository root:

npm run codegen:python
uv sync --project packages/modelparams-python --extra dev
uv run --project packages/modelparams-python pytest packages/modelparams-python/tests

Generated catalog and type files are committed and verified in CI. Python releases use independent modelparams-py@x.y.z tags and publish to PyPI through the repository's trusted-publisher workflow.