SemantaAI Dataset Standard
This document defines the mandatory dataset standard for SemantaAI datasets published under Grencape on Hugging Face.
1) Repository Naming
- Namespace:
Grencape - Repo format:
semantaai-<domain>-<scope> - Examples:
Grencape/semantaai-main-assetsGrencape/semantaai-synthetic-ruGrencape/semantaai-crypto-gold
2) Canonical Layout (Required)
raw/- Immutable source history (OHLCV or source-native records)
- Per-file metadata (
*.meta.json) - Required indexes:
raw/DATASET_INDEX.jsonraw/DATASET_INDEX.csv
- Required docs:
raw/README.md
gold/- Training-ready enriched layer (
raw+ engineered features + targets + labels) - Required indexes:
gold/DATASET_INDEX.jsongold/DATASET_INDEX.csvgold/DATASET_INDEX_FILES.jsongold/DATASET_INDEX_FILES.csv
- Required docs:
gold/README.mdgold/METRICS.mdgold/BUILD_REPORT.json
- Training-ready enriched layer (
3) Naming Rules
- Do not use version folders in public paths (
v1,v2,labeled_v*). - Do not use ambiguous folder names like
main,final,new. - Use semantic layer names only:
raw,gold. - Partitioned gold files must follow:
gold/files/<SYMBOL>__<EXCHANGE>__<TF>__gold.parquet
4) Versioning Policy
- Versioning is metadata-based, not folder-name-based.
- Track iterations in:
gold/BUILD_REPORT.json(built_at_utc, source snapshot, feature set)- git/HF commit history
- optional tags/releases
5) Compatibility Policy
raw/should remain append-only where possible.gold/may evolve, but schema changes must be reflected in:gold/METRICS.mdgold/README.md- index files
6) Minimum Quality Gates Before Publish
- Indexes present and readable (
json+csv) - Metric glossary updated (
gold/METRICS.md) - Build provenance written (
gold/BUILD_REPORT.json) - Root
README.mdlinks to this standard
7) Mandatory Rule For Future Datasets
Every new SemantaAI dataset in Grencape must be published in this standard by default.
Any exception must be temporary and resolved to raw/gold structure before production use.