nima1/decisioncraft-data

Decisioncraft data

This package contains the frozen, locally verified Decisioncraft corpus and its four partitions. Use the full 40-character snapshot commit from Files and versions / history for immutable downloads. The study and independent label review are complete; local verification alone is not a claim that a fresh Hub download was tested.

Purpose and labels

The corpus studies whether a small model can route English developer requests using their explicit context. suggest means a sufficiently specified GNU file/text or local Git request. clarify means a supported operation lacks material information. unsupported means at least one positively requested action is outside the domain. Network, cloud, container, package/runtime, service and database operations are outside scope. A prohibited or quoted operation is not a positive request.

The labels do not authorize shell execution, establish command correctness or assess safety. The project never runs commands from these examples. A separate uncertainty/defer policy must not be conflated with the semantic clarify label.

Construction and provenance

There are 1,800 synthetic rows from 150 scenario groups, not 1,800 independent human-authored requests. An AI coding agent authored 100 local-operation scenarios and 50 external-workflow scenarios, plus a deterministic renderer. Each scenario produces 12 related views: complete/missing or conflicting information, paraphrasing, context placement, recoverable typos, negation and keyword distractions. Unsupported scenarios also include standalone requests and local/external mixtures.

The corpus is original project authoring; it does not import a public corpus or reuse Bashcraft's inspected benchmark or Decisioncraft's M1 development cases. Authoring and rendering are disclosed in the shipped source specifications and builder. Each row retains scenario lineage, variant, rationale, tags, source author and review provenance.

A separate AI reviewer verified all 150 source scenarios, all 1,800 generated derivations and all 360 final cases, binding the review to exact hashes. One duplicate semantic Git scenario was replaced before freezing. Independent-agent review is not human review. The user approved 12 concrete policy examples outside every dataset split and explicitly authorized this disclosed synthetic study. The dataset must not be described as fully human-labeled or human-validated.

Frozen partitions

Partition	Groups	Rows	Rows per label
Train	84	1,008	336
Validation	21	252	84
Calibration	15	180	60
Test	30	360	120

All views of one scenario remain in one partition. Splits use seed 20261004 and the recorded group allocation. Validation selects model/prompt/training settings; calibration fits probabilities and routing thresholds after selection; final test evaluates the frozen protocol.

The test has two cohorts, each with 180 rows from 15 groups and 60 rows per label. The familiar cohort contains new scenarios from represented families. The generalization cohort withholds the complete archive_compression and database_operations families from training, validation and calibration. Report both cohorts; two withheld families cannot establish universal generalization.

The authored expansion has SHA-256 ac346712dbfda1cd5a7cc9fc3cd43d14d78c378d4fb2c9ba24f3a006f0f07b28. The frozen local manifest has SHA-256 2e22842b5d9b34eea8209d19aeb8295376122322ad9638346e02af3b53786db4. These verified content identities identify the exact partitions and manifest distributed here.

Schema and correct use

Rows contain id, group_id, family, request, context, label, rationale, tags, provenance and review. Only request and context are model inputs. Gold labels, family/variant identifiers, rationales and review metadata must stay out of prompts and learned features. The initial per-row review state is preserved; a separate hash-bound ledger records independent review of the frozen corpus.

Once final targets or model errors are used for tuning, the same set is no longer an untouched holdout for those choices. New experiments require a new appropriate holdout. Keep source-group dependence in uncertainty estimates rather than treating every rendered view as independent.

Measured study and exposed limitation

The frozen primary comparison recovered 120/120 ambiguous final cases with adapted Kai versus 62/120 with unchanged Kai, with 0/120 versus 3/120 unnecessary clarifications. Adapted routing covered 332/360 cases; raw argmax classified all 360 correctly. The paired whole-scenario 95% interval for recall improvement was +40.2 to +56.8 percentage points. These results describe this controlled synthetic corpus, not natural-user reliability.

A separate demonstration of all 12 previously approved policy examples produced eight expected routes, three deferrals and one confident mistake: a request to find the literal text docker run in a local note, explicitly without launching Docker, was marked unsupported. Those known policy examples are illustrative rather than an untouched benchmark. This failure exposes a limitation that the perfect synthetic raw score did not reveal. The final data and model were not changed to fix it.

The primary adapted model's temperature scaling also worsened final log loss despite improving calibration-split fitting. The source archive and evidence archive preserve the full report, predictions, numerical analysis, human policy approval and independent reviews. Final labels have now been inspected for reporting; new tuning choices require a new appropriate holdout.

Limits and availability

This is a deliberately balanced, controlled synthetic study, not a sample of natural request frequency. Shared rendering structures and explicit missing-information cues can become shortcuts. Group isolation and exact/delexicalized duplicate checks reduce some leakage but cannot prove absence of all semantic overlap or template effects. Real user behavior, longer contexts and other domains may differ substantially.

The package includes all four JSONL files, manifest, label policy, authored sources, builder, independent review records, the 12-example human policy-review record and license notices. The source archive includes tutorials and software so it remains usable without the private GitHub repository.

Download and inspect

Replace the revision placeholder with this dataset repository's full 40-character commit from Files/history. The dataset revision is distinct from the model repository's revision.

DATASET_REVISION=REPLACE_WITH_FULL_40_CHARACTER_DATASET_COMMIT
uvx --from huggingface-hub==1.33.0 hf download \
  nima1/decisioncraft-data \
  --repo-type dataset \
  --revision "$DATASET_REVISION" \
  --local-dir decisioncraft-data

JSONL can be loaded with the Python standard library; no model or datasets dependency is required. This snippet verifies the public inventory and checks split counts without printing final examples:

python3 - <<'PY'
import hashlib
import json
from pathlib import Path

root = Path("decisioncraft-data")
inventory = json.loads((root / "release-inventory.json").read_text())
for name, expected_hash in inventory["files_sha256"].items():
    assert hashlib.sha256((root / name).read_bytes()).hexdigest() == expected_hash
for split, expected_count in {"train": 1008, "validation": 252,
                              "calibration": 180, "test": 360}.items():
    rows = [json.loads(line) for line in (root / f"{split}.jsonl").read_text().splitlines()]
    assert len(rows) == expected_count
    print(split, len(rows), "rows", len({row["group_id"] for row in rows}), "groups")
PY

The card metadata maps all four files explicitly, including the separately named calibration split. The frozen manifest binds group assignments and source identities. For new training, use only request/context from the appropriate partition, and preserve the split boundaries. Reading final targets to guide a new choice makes this final set unsuitable for an untouched claim about that choice.

License

The original authored specifications, generated synthetic rows and accompanying project documentation use Apache-2.0. Full LICENSE and NOTICE files accompany the package. This license statement does not claim rights to unrelated upstream model training datasets or make this corpus human-authored.