johnahn/mdk-mining-controller-data

MDK Mining Controller — Data

Companion dataset for the mdk-mining-controller prototype (3-week Tether MDK assignment).

What's here

File	Size	What it is
raw/mining_telemetry.parquet	150 MB	5.2 M rows of 1-minute telemetry for 30 ASIC miners across 120 days. 21 columns: hashrate, power, voltage, frequency, temperature, ambient temp, operating mode, is_online, hardware_model_id, miner_id, timestamp, plus failure-scenario metadata. Output of src/synthetic/generator.py.
raw/mdk.duckdb	201 MB	Same telemetry as the parquet, loaded into a DuckDB database. Used by the batch pipeline for fast columnar queries. Idempotent re-derivation from the parquet takes ~30 s.
processed/features.v3.parquet	3.6 GB	5.2 M rows × 175 features — the fully engineered feature matrix used to train the XGBoost + LSTM-AE models. Includes the TE KPI and its rolling / trend / correlation variants. Output of src/pipeline/features.py:build_feature_matrix. Rebuilding from raw costs ~25 min.

All files are reproducible from the repo's synthetic generator — nothing here is real mining data. The dataset exists to let reviewers skip the 40-minute rebuild step.

Usage

Clone the code repo first:

git clone https://github.com/john-yo-ahn/mdk-mining-controller
cd mdk-mining-controller
uv sync

Then download this dataset into the expected layout:

uv run python -c "
from huggingface_hub import snapshot_download
snapshot_download(
    'johnahn/mdk-mining-controller-data',
    repo_type='dataset',
    local_dir='data',
)
"

Now the repo has the full data/raw/ + data/processed/ tree. Run:

uv run mdk check      # 13/13 pipeline invariants, ~11 min
uv run mdk validate   # 4 end-to-end tests, ~9 min
uv run mdk            # live Textual dashboard, loads real models

Provenance & seeding

All artifacts were generated deterministically with seed=42 across the generator, split, and model training. Reruns produce byte-identical metrics to the sidecar metadata in the code repo under data/models/*.metadata.json. The mdk check harness explicitly verifies this on every invocation.

Schema

raw/mining_telemetry.parquet (21 columns)

miner_id, timestamp, hashrate_th, power_w, voltage_v, frequency_mhz, temperature_c, ambient_temperature_c, operating_mode, is_online, hardware_model_id, hardware_model, hash_board_serial, scenario_name, scenario_onset_step, scenario_duration, is_pre_failure, fan_rpm, error_count, voltage_sag, hashrate_error.

processed/features.v3.parquet (175 columns)

Derived from the raw 21 + intermediate 34 columns. Categories:

• Ratios (~10): efficiency_jth, temp_delta_c, power_per_ghz, voltage_deviation, hashrate_realization, te_base, te_adjusted, te_health, …
• Rolling statistics (~80): {metric}_roll_{60|360|10080}m_{mean|std|min|max} for all 10 base signals
• Trend features (~20): linear-regression slopes and rate-of-change over 60/360-minute windows
• Cross-signal correlations (~15): voltage-temperature, power-hashrate, TE-voltage, etc. rolling correlations
• Diurnal features (~6): hour-of-day, day-of-week, sine/cosine encodings
• Cross-miner features (~10): container-level means and deviations from container baseline
• Labels: is_pre_failure (binary target), failure_type (multi-class for analysis)

Full schema enumerable via build_feature_matrix in src/pipeline/features.py in the code repo.

License

MIT. Same as the code repo.

Citation

Not peer-reviewed work — a prototype built against Tether's MDK assignment spec. If you reference it, a link back to both repos is appreciated:

• Code: https://github.com/john-yo-ahn/mdk-mining-controller
• Data: https://huggingface.co/datasets/johnahn/mdk-mining-controller-data