sewlore/synthetic-fabric-csv-validation-cases

Sewlore Synthetic Fabric CSV Validation Cases

All 72 records are invented software inputs. This is a small, deterministic CSV-validation teaching corpus. It contains no physical fabric measurements, garment trials, personal data, photographs or washing observations. No model was trained or evaluated. AI assistance was used to prepare cases and documentation; labels were captured by executing a frozen public Python package.

Use it to learn a specific parser contract, compare regression behavior or build a validation exercise. Expected labels describe sewlore-fabric-care==0.1.0; they are not independent scientific ground truth, a textile dataset or a general benchmark of validators.

Read the dataset documentation and fixed-version downloads for a literal preview, the complete data dictionary, source checksums and reproduction steps. The 72 cases remain synthetic software inputs.

Quick overview

Property	Value
Creator	Sewlore
Cases	72, each with a stable case-001–case-072 ID
Intended exercises	18 normal, 41 invalid-intent, 13 numeric/lexical edge cases
Independent core results	30 accepted; 42 rejected
Independent strict-policy results	27 accepted; 45 rejected
Whole-corpus core result	30 accepted records; 46 diagnostic entries
Whole-corpus strict result	0 accepted records; 5 policy diagnostic entries
License	MIT; full notice in LICENSE.txt

The exercise category is an intention, not its outcome. Three rows are accepted by the core but rejected by the stricter policy. Submitting all rows together rejects the entire file at the policy stage; per-case results do not describe a whole-file export.

Browse the data

data/cases.csv joins the original input, case catalog, expected labels and diagnostics into one 26-column viewing table. Its one test split denotes software test fixtures, not an ML training/evaluation split. No examples were held out. Every field is explicitly loaded as a string so invalid numbers, blank tokens, whitespace and JSON diagnostics retain their literal meaning.

Do not submit this joined table to the validator. Its ten expected input columns are available in the unchanged synthetic input CSV. The data dictionary explains that native schema and labels.

The joined table adds core_diagnostics_json and policy_diagnostics_json, containing JSON arrays of actual version-specific diagnostic objects. Parse those columns with json.loads. Error counts count diagnostic entries, not failed physical samples. All is_synthetic tokens are true.

To explore a local download using only Python's standard library:

import csv
import json
from collections import Counter

with open("data/cases.csv", encoding="utf-8", newline="") as handle:
    cases = list(csv.DictReader(handle))
assert len(cases) == 72
assert Counter(row["core_status"] for row in cases) == {"accepted": 30, "rejected": 42}
policy_difference = [row for row in cases
                     if row["core_status"] == "accepted" and row["policy_status"] == "rejected"]
assert len(policy_difference) == 3
for row in policy_difference:
    print(row["case_id"], row["feature"], json.loads(row["policy_diagnostics_json"]))

For Hugging Face Datasets users, the local copy can be loaded with the same raw string schema:

import csv
from datasets import Features, Value, load_dataset

with open("data/cases.csv", encoding="utf-8", newline="") as handle:
    columns = next(csv.reader(handle))
fixtures = load_dataset(
    "csv", data_files={"test": "data/cases.csv"},
    features=Features({name: Value("string") for name in columns}),
    keep_default_na=False, na_filter=False,
)["test"]
assert len(fixtures) == 72

Files and exact reproduction

• data/synthetic_cases.csv: unchanged ten-column raw inputs.
• data/case_catalog.csv: IDs, logical record mapping and intended features.
• data/expected_labels.csv: separate independent per-case labels and core values.
• data/expected_diagnostics.json: complete per-case core and policy diagnostics.
• data/expected_batch_results.json: the distinct whole-corpus result.
• data/cases.csv: deterministic convenience join for the Dataset Viewer.
• generate.py, verify_preparation.py: original frozen-wheel generator and checks.
• build_viewer_table.py: join builder; it performs no validation or relabelling.
• PROVENANCE.json, DATA_DICTIONARY.md, LICENSE.txt: provenance, schema and license.

Python 3.10 or newer is required for the reproduction scripts. In a local copy:

python -m pip download --only-binary=:all: --no-deps sewlore-fabric-care==0.1.0 --dest wheels
python generate.py --wheel wheels/sewlore_fabric_care-0.1.0-py3-none-any.whl --check
python verify_preparation.py --wheel wheels/sewlore_fabric_care-0.1.0-py3-none-any.whl
python build_viewer_table.py --check

The first command downloads the public wheel. The scripts themselves make no network requests or account changes. The required wheel SHA-256 is:

3fda5e77c92aea891b5c4ea1f53f0bfac6f1d7bc1a3ff0003e725bf70f6f0606

Labels come from actual validate_csv(text) and validate_submission(bytes) outputs, with each case evaluated alone. The core accepts independently valid rows. The strict policy requires bounded UTF-8 input, anonymous sample- IDs, enumerated care codes and empty notes; any policy violation rejects that file before calculation. The fixed source revision and public package 0.1.0 make the software contract inspectable.

Scope and limitations

Cases are explicitly selected lexical fixtures, not random or representative observations. Care-method strings test syntax; no corresponding treatment took place. Deliberate tokens include zero, negative, missing, NaN, unit text, scientific notation, unsupported units and overflow/underflow combinations. The CSV itself is structurally well formed. Malformed quoting, wrong headers, byte-limit and row-limit cases are outside this corpus.

Computed percentages describe hypothetical arithmetic. Positive means contraction; negative means expansion. Area uses a rectangular assumption only. Output tokens preserve 17 significant digits of the frozen binary float; this implies no physical measurement precision. Different numerical implementations should compare with a documented tolerance.

This resource supports software learning and regression analysis. It cannot establish textile suitability, care advice, garment fit, empirical error rates, model quality or results of a physical experiment. No academic paper, peer review, institutional affiliation or individual researcher contribution is claimed.

Related learning resources

The Sewlore fabric measurement methods guide explains record-keeping context. It is a contextual guide, not the origin of real observations in this corpus. The public notebook lesson shows how to review accepted records and corrections. Package API documentation explains the generic and strict entry points.

Code, original case definitions and documentation: © 2026 Sewlore, MIT. The full existing license notice is retained unchanged. This license grants no trademark ownership or endorsement.