eval-dataset-smith-py 0.1.0

Generate balanced AI eval fixtures from source examples, bugs, docs, and policies. Python port of @mukundakatta/eval-dataset-smith.

eval-dataset-smith-py

Generate balanced AI eval fixtures from your bugs, docs, examples, and policies. Zero runtime dependencies.

Python port of @mukundakatta/eval-dataset-smith. The JS sibling has the full design notes; this README sticks to the Python API.

Install

pip install eval-dataset-smith-py

Usage

from eval_dataset_smith import forge_dataset, stratified_split

sources = [
    {"type": "bug",  "id": "B-1", "input": "repro: click X",         "expected": "no crash",        "difficulty": "easy"},
    {"type": "bug",  "id": "B-2", "input": "repro: open file Y",     "expected": "no crash",        "difficulty": "med"},
    {"type": "doc",                "question": "how does foo work?", "answer": "see chapter 3",     "difficulty": "easy"},
    {"type": "policy",             "input": "is PII allowed?",       "expected": "redact",          "difficulty": "hard"},
]

ds = forge_dataset(sources, balance_keys=["type", "difficulty"])

ds.cases       # list[EvalCase]   -- the eval fixtures
ds.balance     # {"type": {...}, "difficulty": {...}} -- audit input skew
len(ds)        # 4

# Per-tag stratified split (preserves type balance across train/test)
parts = stratified_split([c.__dict__ for c in ds.cases], ratio=0.8)
parts["train"], parts["test"]

API

forge_dataset(sources, balance_keys=("type","difficulty"), max_per_type=20) -> Dataset

Top-level Pythonic entry point. Returns a typed Dataset of EvalCase records plus a balance histogram you can use to audit input skew.

build_eval_dataset(items, max_per_type=20) -> list[dict]

Direct port of the JS buildEvalDataset. Accepts the JS field-name aliases:

Field	Aliases
input	input / question / prompt
expected	expected / answer / acceptance
type	type (defaults to "general")
tags	tags: list[str]

stratified_split(items, ratio=0.8) -> {"train": [...], "test": [...]}

Direct port of the JS stratifiedSplit. Splits by the first tag of each item, slicing each group at ceil(len(group) * ratio).

API differences from the JS sibling

• forge_dataset is a Python addition that returns typed dataclasses (Dataset, EvalCase).
• build_eval_dataset and stratified_split mirror the JS function names with snake_case.

See the JS sibling's README for the full design notes.