README configs/dataset_info regenerated from the actual Parquet (all
nine configs loaded and split-counts cross-checked, 9/9 OK; features
round-tripped through Features._from_yaml_list). Adds the 2026-08-16
changelog entry, the serving-format note in the v2 section, and two new
hard rules in CLAUDE.md: one data format across ALL configs (the hub
infers a single builder from the first config's files), and dataset_info
dtypes must be YAML-format strings. dataset_card.json updated in sync;
.agents/.claude tooling caches ignored.
Claude-Session: https://claude.ai/code/session_0188gRcGZdDN828oyp4TGkt7
Co-Authored-By: Claude Fable 5 (max) <noreply@anthropic.com>
data: serve all configs from Parquet (add v2_parquet conversions)
Adds v2_parquet/<config>/train-00000.parquet for the four v2 configs,
built and fidelity-verified by spikenaut-etl build-v3 (ETL merge
611066b, PR #9): identical row counts, rows, columns, and nulls versus
the JSONL, with one representation change -- qubic_ticks.timestamp is
timestamp[ms] instead of timestamp[s] (Parquet has no seconds
resolution; values unchanged). full_data/*.jsonl are byte-untouched and
remain the canonical cleaned exports.
Why: a script-less Hub dataset resolves ONE packaged builder from the
first config's data files and applies it to every config, so the
2026-08-15 card (JSONL v2 + Parquet v3) made the viewer parse Parquet
bytes with the JSON loader and every v3 config failed first-rows.
Claude-Session: https://claude.ai/code/session_0188gRcGZdDN828oyp4TGkt7
Co-Authored-By: Claude Fable 5 (max) <noreply@anthropic.com>
fix: serialize dataset_info dtypes in the card's YAML format
The previous card wrote feature dtypes as repr strings
("Value('float64')"), which the Hub's config-names step cannot parse,
breaking the whole Dataset Viewer. dtypes now come from
Features._to_yaml_list() - the serializer datasets itself uses - and the
generator round-trips every features block through
Features._from_yaml_list() (the exact parser the viewer runs) before the
card is written, so this class of breakage fails at build time.
Claude-Session: https://claude.ai/code/session_0188gRcGZdDN828oyp4TGkt7
Co-Authored-By: Claude Fable 5 (max) <noreply@anthropic.com>
docs: rewrite dataset card for v3 action-proposal restructure
Retags the dataset reinforcement-learning (time-series-forecasting kept
for the v2 configs) and adds the v3 sections: Control Hierarchy & Safety
(deterministic filter is final authority; shadow/assist/active; filter-
override rate as the shadow KPI), Telemetry Contract (schema_version
semver, sampling-rate table, units, Q8.8 convention), Action Taxonomy
(frozen ids + teacher rules with thresholds and rule ids), Encoding
(raw + params policy, axon-encoder semantics, dt_seconds is prescriptive),
Episode & Split Design (leakage rules), and Migration/Changelog.
configs and dataset_info are generated from the actual data files; every
listed config was loaded with datasets.load_dataset and its split sizes
cross-checked before this commit (9/9 OK). The empty action_proposals /
safety_filter_log artifacts are documented but not listed as configs:
datasets <= 5.0.x cannot batch a 0-row Parquet file. The frontmatter
'metrics: accuracy' key is removed - it referred to the disavowed 95.2%
figure.…
data: add v3 action-proposal trajectory tree (schema 3.0.0)
Additive v3/ Parquet tree built from the published v2 JSONL by
spikenaut-etl build-v3 (rmems/spikenaut-telemetry-etl 983323b, teacher
policy v1.0.0). v2 files are untouched.
- state_telemetry (52 cols) + outcomes: 569,344 / 118,784 / 117,653
rows across chronological train/validation/test blocks; episodes are
fixed 4,096-step windows; embargo episodes gpu-000139 and gpu-000169
are published in no split (gap >= outcome horizon H=64).
- gpu_telemetry_v3 + qubic_signals: 813,973-row flat derivations under
a single 'full' split; qubic_* columns separated from GPU sensors,
duplicate clock_mhz dropped after byte-equality verification.
- encoding_params: 36-feature axon-encoder sidecar, min/max fitted on
the train split only.
- action_proposals / safety_filter_log: typed empty schema artifacts.
The teacher emitted zero labels over the v2 backfill because the core
signals (NVML throttle mask, ECC counters) were never collected, and
it does not guess.
- ts_utc / ts_synthetic_…
node_sync_harvest.jsonl ended with 12 rows that are not capture data:
appended ~21 hours before the record preceding them, the only rows in the
file carrying a UTC offset, and holding two distinct
(power_w, gpu_temp_c) pairs between them -- (0.0, 80.0) and (400.0, 40.0).
They survived the 2026-08-03 rebuild because that release restored all 20
rows the old pipeline's tail-trim heuristic had removed. The heuristic
fired for the wrong reason -- on its own corrupted all-zero output -- but
it was not wrong that some tail rows are synthetic. Restoring all 20
over-corrected. beads sv-l43 flagged these in July as placeholders that
teach the SNN an idle pattern that never occurred.
No gate caught it. The fabrication check looks for uniform spacing, and
12 rows in 120,334 move no distribution-based test. The pipeline now
quarantines a short out-of-order run at the end of a file and fails the
build on any time reversal it did not quarantine. The trigger is the
ordering break, not the constant …
Records what is not discoverable from the files themselves: that this is
a published HF artifact rather than a source repo, that every JSONL is
generated by rmems/spikenaut-telemetry-etl and must not be hand-edited,
that the irreplaceable pre-cleaning originals are untracked and live only
in ~/spikenaut-data-backup and Spikenaut-Vault/data, and the two viewer
config rules learned from breaking it (one config per file; never list a
subset of another file in the same config).
Also documents clean_datasets.jl as superseded and destructive, and the
standing rules that came out of it: never synthesize a timestamp, missing
means null rather than "" or 0.0, and assert output is non-degenerate
before writing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Restores the 934,287 records that prior revisions advertised but did not
contain, and adds chain attribution the old pipeline had discarded.
neuromorphic_data.jsonl 813,973 identical {"telemetry":{}} -> 813,973 rows, 12 sensor columns
node_sync_harvest.jsonl 120,314 rows, every numeric 0.0 -> 120,334 rows, real values
Both were rebuilt from the pre-cleaning originals (395 MB / 57 MB) through
rmems/spikenaut-telemetry-etl, which gates every release on: no constant
or all-null columns, a minimum distinct-row ratio, bounded row-count
drift, non-fabricated timestamps, and an exact schema match. Round-trip
verified against the originals, zero mismatches.
Also in this release:
- Real timestamps restored. The prior script overwrote 114,250 of them
with a generated 10s ramp because its ISO-8601 check tested for a
literal "T" while the source used a space separator. The mining window
is 2026-03-19 11:55 -> 03-20 14:05, not the advertised span through
April 2, which never …
Correct dataset card: 98% of advertised records are content-free
An audit found that 934,287 of the 953,290 advertised records carry no
information:
neuromorphic_data.jsonl 813,973 lines, 1 distinct line: {"telemetry":{}}
node_sync_harvest.jsonl 120,314 lines, all numerics 0.0, blockchain ""
Root cause is a key-type bug in the cleaning script that produced them:
JSON was read into a Symbol-keyed dict but filtered and looked up with
String keys, so every lookup silently returned its default and every
filter matched nothing. The same script overwrote 114,250 real timestamps
with a fabricated evenly-spaced sequence, because its ISO-8601 check
tested for a literal "T" and the real timestamps used a space separator.
Card corrections:
- Add dated Data Quality Notice naming both corrupt files
- Restate record count: 953,290+ -> 60,003 usable
- Remove the per-coin attribution table; the data cannot support it
(only dynex 5,001 and qubic 1,083 are labelled; 114,250 are not)
- Correct the collection window: three disjoint March 2026 windows,
not …
Flatten nested JSONL objects for HF Dataset Viewer compatibility
fix: Correct YAML metadata for HuggingFace validation
Raul MC
da598e8
- Use 'en' for language (ISO 639-1 code)
- Use 'monolingual' for multilinguality (string format)
- Use official 'univariate-time-series-forecasting' task ID
- Fix validation errors and enable successful push
fix: Add YAML metadata frontmatter to dataset card
