pfnstudio-core

The Python contract for PFN Studio FM projects.

from pfnstudio_core import Prior, Model, Run, register_block, register_prior, register_scorer
from pfnstudio_core.scorers.base import DatasetScorer, ScorerResult

@register_prior("my_prior")
class MyPrior(Prior):
    def sample(self, seed: int): ...

@register_block("my_attention")
class MyAttention:
    def __init__(self, d_model: int, n_heads: int): ...

# Paper-specific scorer — ships in the template beside evals/<slug>.yaml.
@register_scorer("my_eval")
class MyScorer(DatasetScorer):
    def score(self, *, model, eval_spec, loader, run_spec) -> ScorerResult: ...

The CLI discovers anything registered via these decorators and validates models/*.yaml references against the registry.

Layout

• prior.py — Prior ABC and built-in prior loader
• model.py — Model config + block-composition
• eval.py — EvalSpec — the declarative benchmark spec (dataset + metrics + baselines)
• scorers/ — DatasetScorer — the executable scoring pipeline; core ships only generic scorers, paper-specific ones live in templates
• run.py — Run manifest + executor protocol
• registry.py — @register_prior, @register_block, @register_scorer and discovery
• loaders.py — load YAML artifacts into typed objects
• blocks/ — built-in architecture blocks (transformer encoder, causal attention, heads)
• training/ — minimal in-process training loop for the local compute adapter