Datasets:

Crovia
/

global-ai-training-omissions

Error code:   FeaturesError
Exception:    ArrowInvalid
Message:      Schema at index 1 was different: 
schema: string
generated_at: double
context: struct<what_this_is: string, what_this_is_not: string>
start_here: struct<ranking_jsonl: string, canon: string, snapshot: string>
top_omissions: list<item: struct<necessity_id: string, score: double, avg_persistence_days: double, records: int64>>
vs
last_sync: string
sync_method: string
components: struct<badges: struct<count: int64, last_update: string, index_file: string>, cards: struct<count: int64, last_update: string, directory: string>>
coherence_checks: struct<badges_cards_match: bool, cards_count: int64, badges_count: int64, last_check: string>
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 243, in compute_first_rows_from_streaming_response
                  iterable_dataset = iterable_dataset._resolve_features()
                                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 3608, in _resolve_features
                  features = _infer_features_from_batch(self.with_format(None)._head())
                                                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2368, in _head
                  return next(iter(self.iter(batch_size=n)))
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2573, in iter
                  for key, example in iterator:
                                      ^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2060, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2082, in _iter_arrow
                  yield from self.ex_iterable._iter_arrow()
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 572, in _iter_arrow
                  yield new_key, pa.Table.from_batches(chunks_buffer)
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "pyarrow/table.pxi", line 5039, in pyarrow.lib.Table.from_batches
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
              pyarrow.lib.ArrowInvalid: Schema at index 1 was different: 
              schema: string
              generated_at: double
              context: struct<what_this_is: string, what_this_is_not: string>
              start_here: struct<ranking_jsonl: string, canon: string, snapshot: string>
              top_omissions: list<item: struct<necessity_id: string, score: double, avg_persistence_days: double, records: int64>>
              vs
              last_sync: string
              sync_method: string
              components: struct<badges: struct<count: int64, last_update: string, index_file: string>, cards: struct<count: int64, last_update: string, directory: string>>
              coherence_checks: struct<badges_cards_match: bool, cards_count: int64, badges_count: int64, last_check: string>

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

 ██████╗██████╗  ██████╗ ██╗   ██╗██╗ █████╗ 
██╔════╝██╔══██╗██╔═══██╗██║   ██║██║██╔══██╗
██║     ██████╔╝██║   ██║██║   ██║██║███████║
██║     ██╔══██╗██║   ██║╚██╗ ██╔╝██║██╔══██║
╚██████╗██║  ██║╚██████╔╝ ╚████╔╝ ██║██║  ██║
 ╚═════╝╚═╝  ╚═╝ ╚═════╝   ╚═══╝  ╚═╝╚═╝  ╚═╝

📊 GLOBAL AI TRAINING OMISSIONS

Append-Only Temporal Observation Ledger

⚓ Registry Observer · � Dataset Files · 📖 Evidence Format

⚠️ WHAT WE TRACK

┌────────────────────────────────────────────────────────────────────────────┐
│                                                                            │
│   This dataset records OBSERVATIONS — not analysis or judgment.            │
│                                                                            │
│   📊 Evidence files:         Verifiable, timestamped records               │
│   📜 Canon definitions:      NEC# vocabulary (necessities.v1.yaml)         │
│   🎯 Observation types:      Presence, absence, temporal pressure          │
│   👤 Public ledger:          Append-only, cryptographically anchored       │
│                                                                            │
│   The data is PUBLIC. The method is DETERMINISTIC.                         │
│   This is PROTOCOL v1.0 — Founding Stage, Jan 2026.                        │
│                                                                            │
└────────────────────────────────────────────────────────────────────────────┘

📖 What This Dataset Is

This dataset is an append-only mirror of Crovia's public Training Provenance Registry (TPR).

What This Dataset Contains

✅ Observed presence/absence events — timestamped facts from the registry
✅ Temporal metrics — first_seen, last_seen, days_monitored (derived mathematically)
✅ Cryptographic receipts — receipt_hash for each observation
✅ Merkle-root anchored history — verifiable integrity proofs
✅ Registry metadata — source, endpoint, timestamps

What This Dataset Does NOT Contain

❌ Scores — no shadow scores, trust scores, or compliance scores
❌ Rankings — no leaderboards or comparative judgments
❌ Badges — no GOLD/SILVER/BRONZE classifications
❌ Compliance judgments — no violation assessments or interpretations
❌ Placeholders — every field is derived from real observations

This dataset is a temporal evidence ledger, not an analysis or ranking system.

🔬 How Observations Work

┌─────────────────────────────────────────────────────────────────┐
│                    OBSERVATION PIPELINE                         │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  1. OBSERVE           2. RECORD            3. PUBLISH           │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐      │
│  │ Scanner      │───▶│ TPR Registry │───▶│ Evidence     │      │
│  │ (automated)  │    │ (PostgreSQL) │    │ Dataset      │      │
│  │              │    │ Append-only  │    │ (this repo)  │      │
│  └──────────────┘    └──────────────┘    └──────────────┘      │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

📁 Dataset Structure

├── observations.jsonl         # PRIMARY: Append-only observation ledger (213 records)
├── EVIDENCE.json              # Legacy: Static evidence file
├── badges/                    # Legacy: Deprecated (not used)
├── cards/                     # Legacy: Deprecated (not used)
├── canon/
│   └── necessities.v1.yaml    # NEC# definitions (20 types)
├── open/
│   ├── forensic/              # Evidence processing scripts
│   ├── signal/                # Presence/absence signals
│   └── temporal/              # Historical pressure data
└── v0.1/                      # Versioned snapshots

Primary File: `observations.jsonl`

Each line is a JSON record with:

receipt_hash: Cryptographic identifier (sha256)
target_id: Model/dataset identifier (e.g., openai/whisper-large-v3)
observation_type: presence or absence
observed_at: Timestamp (ISO 8601)
first_seen, last_seen: Temporal boundaries
days_monitored: Days between first and last observation
observation_count: Total observations for this target
absence_streak_days: Consecutive absence days (if applicable)
source: Observer identifier
registry_endpoint: Source API
merkle_root: Registry integrity proof

⚓ Registry Observer

View Live Registry →

The Registry Observer Space displays:

Registry statistics (total observations, today, unique targets)
Merkle root (cryptographic verification)
Recent observations (timestamped records)
Truth Anchor (source-of-truth declarations)

Note: The Observer is viewer-only. It does NOT calculate scores, assign badges, or perform analysis.

🔗 Using This Dataset

For Researchers

Access evidence files directly:

# Download evidence file
wget https://huggingface.co/datasets/Crovia/global-ai-training-omissions/resolve/main/EVIDENCE.json

# View canon definitions
wget https://huggingface.co/datasets/Crovia/global-ai-training-omissions/resolve/main/canon/necessities.v1.yaml

For Verification

# Verify snapshot integrity
curl -s https://huggingface.co/datasets/Crovia/global-ai-training-omissions/resolve/main/snapshot_latest.json | jq '.hash'

What this dataset is

Crovia — Global AI Training Omissions Evidence Dataset v0.1

This dataset publishes verifiable, hash-anchored evidence of observations recorded across AI training datasets and models.

It answers one and only one question:

Are public AI training disclosures observable — yes or no?

This dataset IS

an observation layer, not an audit
a cryptographically verifiable record
a public, reproducible signal of absence or presence
aligned with EU AI Act transparency principles

This dataset IS NOT

it does not audit models
it does not infer intent
it does not assign blame
it does not make legal claims

Observable, verifiable updates

This dataset is designed for automatic updates:

Public artifacts are observed systematically
If nothing changes, the update itself proves persistence
If something changes, hashes and commits reflect it
No manual curation or interpretation
No retroactive edits

Current Status (Jan 2026): Founding Stage

5 curated models demonstrate protocol feasibility
Automation infrastructure ready for deployment
Full-scale observation pending infrastructure activation

Every update will be:

Publicly committed to this dataset
Reproducible via open scripts (open/forensic/)
Independently verifiable via cryptographic anchors

Start here (viewer-first)

If you open only one file, open:

➡️ START_HERE.md

It explains the evidence layout for non-technical readers.

Open Plane (public observation layer)

The Open Plane measures one condition only:

Absence is observable.

It contains:

presence signals:
open/signal/presence_latest.jsonl
absence receipts (time-bucketed):
open/forensic/absence_receipts_7d.jsonl
overview:
open/README.md

Core artifacts

observation records: global_ranking.jsonl (legacy filename, contains observations)
current snapshot: snapshot_latest.json
cryptographic proof: EVIDENCE.json
canonical vocabulary: canon/necessities.v1.yaml

PRO Shadow (non-disclosing)

Crovia PRO can compute private semantic measurements.

The Open Plane publishes a hash-anchored shadow pointer proving that a measurement exists without disclosing private data:

open/signal/pro_shadow_pressure_latest.json
open/README_PRO_SHADOW.md

Temporal pressure (silence over time)

Crovia tracks how long silence persists under sustained observation.

Temporal pressure increases when:

observation coverage is HIGH
no public training evidence is disclosed
silence persists across days

This does not imply wrongdoing.

➡️ open/temporal/temporal_pressure_30d.jsonl

Downloads last month: 1,294