MacroLens / code /dataloader /__init__.py
itouchz's picture
Add code/macrolens/ + code/dataloader/ packages; fix README path prefixes
f02626b verified
Raw
History Blame Contribute Delete
1.24 kB
"""MacroLens dataloader: canonical sample budgets + index generators.
Sits between the immutable benchmark artifacts (`data_small_caps/benchmark/`)
and the family runners (`baselines/`). Owns:
- `budgets.EVAL_N_PER_TASK`, `budgets.TRAIN_N_PER_TASK`, `budgets.SEED`:
the canonical sample budgets per task. Single source of truth.
- `canonical_indices.get_canonical_indices(task, split)`: deterministic,
stratified index generator with on-disk cache. Same indices for every
method, so cross-method comparison on each task is fair.
Cache layout: `experiments/cache/canonical_indices/<key>/<split>_<task>.parquet`
where `<key>` encodes (n_eval, n_train, seed, stratifier_version).
Changing any cache-key dimension produces a new cache directory; the
immutable benchmark artifacts under `data_small_caps/` are NEVER touched
(the cache lives experiment-side, not in the dataset tree).
"""
from .budgets import (
EVAL_N_PER_TASK,
TRAIN_N_PER_TASK,
SEED,
STRATIFIER_VERSION,
cache_key,
)
from .canonical_indices import get_canonical_indices, build_all
__all__ = [
"EVAL_N_PER_TASK",
"TRAIN_N_PER_TASK",
"SEED",
"STRATIFIER_VERSION",
"cache_key",
"get_canonical_indices",
"build_all",
]