cep-capsules / README.md
CroviaTrust
Update README.md
0e3dcd8 verified
# 🟣 Crovia — CEP Capsules (v1)
**Crypto Evidence Packages for the AI Act Era**
Portable, offline-verifiable provenance capsules for the world’s most-used open datasets.
> **If a dataset shaped modern AI, it deserves a receipt.**
> These are the first publicly verifiable evidence capsules of their kind.
---
# 🚀 What Are CEP Capsules?
A **CEP.v1 capsule** (*Crypto Evidence Package*) is a compact, self-contained file that proves:
- **where attribution signals came from**
- **how payouts would be computed**
- **which trust bundle verified the run**
- **what the hashchain root is**
- **that everything inside is immutable**
Each capsule is:
- **📦 portable** (just a JSON file)
- **🔍 independently verifiable** (3 lines of Python)
- **🛡 AI Act–aligned** (trust bundle, receipts, payouts)
- **⛓ backed by a hashchain** (tamper-evident)
Think of this not as a dataset —
**but as the evidence layer underneath datasets.**
---
# 📘 Included Capsules (Dec 2025 Preview)
| Dataset Slice | Capsule |
|---------------|---------|
| C4 | [CEP-C4-2025-12.json](./CEP-C4-2025-12.json) |
| LAION 5B (sample) | [CEP-LAION-2025-12.json](./CEP-LAION-2025-12.json) |
| Wikipedia | [CEP-WIKIPEDIA-2025-12.json](./CEP-WIKIPEDIA-2025-12.json) |
| Wikitext | [CEP-WIKITEXT-2025-12.json](./CEP-WIKITEXT-2025-12.json) |
| FineWeb | [CEP-FINEWEB-2025-12.json](./CEP-FINEWEB-2025-12.json) |
| The Pile | [CEP-PILE-2025-12.json](./CEP-PILE-2025-12.json) |
| ArXiv abstracts | [CEP-ARXIV-2025-12.json](./CEP-ARXIV-2025-12.json) |
| OpenSubtitles | [CEP-OPENSUB-2025-12.json](./CEP-OPENSUB-2025-12.json) |
| Stack / Code | [CEP-STACK-2025-12.json](./CEP-STACK-2025-12.json) |
| BookCorpus | [CEP-BOOKCORPUS-2025-12.json](./CEP-BOOKCORPUS-2025-12.json) |
> These capsules do **not** contain dataset content.
> They contain **evidence about how attribution signals flow through Crovia’s open engine.**
---
# 🧪 Verify Any Capsule in 3 Lines
Save a capsule (e.g. `CEP-C4-2025-12.json`) locally and run:
import json, hashlib
with open("CEP-C4-2025-12.json") as f:
cep = json.load(f)
root = cep["hashchain"]["root"]
print("Root:", root)
print("Valid:", root == hashlib.sha256(cep["payouts"].encode()).hexdigest())
Everything is verifiable **offline**, with **no token**, **no API**, **no Crovia server**.
---
# 🧠 Why This Matters
Modern AI models are trained on massive public datasets…
…but nobody can **prove**:
- what signals came from where
- how much each source contributed
- how payouts would flow
- what trust criteria were applied
- whether logs were tampered with
**Crovia introduces the evidence layer that the ecosystem was missing.**
A standard way to ship **real, inspectable provenance** with AI training pipelines.
These capsules:
- help researchers audit models
- help companies comply with the AI Act
- help dataset creators receive attribution visibility
- help the community trust what models are built on
---
# 👇 Want to Explore or Collaborate?
Crovia is entirely community-driven.
We're looking for:
- dataset maintainers
- compliance researchers
- cryptography engineers
- model evaluators
- people who care about transparent AI
If you want to contribute, explore, or join early pilots:
- **Open an issue on this dataset**
- **Star the dataset to follow updates**
- **Mention @Crovia on LinkedIn or X — we respond to everyone**
---
# 🧩 Roadmap (Public Layer)
- CEP.v2 — multi-run lineage
- DSSE Open Integration (semantic signal explorer)
- Verified Dataset Manifests
- Training Pipeline Attestations
---
# 🙌 Credits
Crovia is an independent initiative committed to
**transparent, evidence-based AI attribution.**
This preview is released under an open license to accelerate adoption
and give researchers the tools missing from the ecosystem.
---
# ⭐ If you find this useful…
**Please star the dataset.**
It helps more researchers discover the project — and it signals that this space matters.
Dataset home:
https://huggingface.co/datasets/Crovia/cep-capsules