rmems/Spikenaut-SNN-Telemetry

🧠 Spikenaut SNN v2 Telemetry Dataset

"The threshold at which stimulus becomes perceptible"

GPU telemetry, blockchain mining telemetry, and HFT paper-trading data for Spikenaut SNN v2.

993,298 records across four sources, collected March 2026 on an NVIDIA RTX 5080 (Blackwell SM_120).

---

📊 Dataset Overview

Config	Records	Window	Description
gpu_telemetry	813,973	—	RTX 5080 sensors: power, temps, clocks, utilization
mining	120,322	2026-03-19 11:55 → 03-20 14:05	Multi-coin node sync telemetry
hft	31,573	2026-03-11 18:22 → 03-12 02:40	Ghost Money paper-trading log
qubic_ticks	27,430	2026-03-20 08:55 → 03-21 08:46 UTC	Qubic tick stream in SNN format

These are disjoint capture windows, not one continuous run. Format is JSONL, one flat record per line.

from datasets import load_dataset

ds = load_dataset("rmems/Spikenaut-SNN-Telemetry", "mining", split="train")

---

📂 Files

Full data

File	Records	Description
full_data/neuromorphic_data.jsonl	813,973	GPU telemetry, 12 sensor columns
full_data/node_sync_harvest.jsonl	120,322	Mining telemetry with chain attribution
full_data/ghost_market_log.jsonl	31,573	HFT paper-trading log
full_data/qubic_ticks_snn.jsonl	27,430	Qubic ticks, SNN input format
full_data/qubic_ticks.jsonl	27,430	Raw Qubic tick stream (source of the above — not additional records)

Samples

samples/{gpu,mining,hft,qubic}_SAMPLE_{100,1k}.jsonl — seeded random draws from the corresponding full file, for quick inspection. They are not prefixes and are excluded from the viewer configs, so loading a sample alongside its parent never double-counts rows.

Model artifacts — ⚠️ provenance unverified

File	Description
full_data/snn_model.json	16-neuron LIF definition. All 16 neurons carry identical weights, 10 of 16 weights are 1.0e-45 denormals, all thresholds 1.0, all decay 0.85.
full_data/hybrid_training_results.json	Reports total_samples: 8 and Kaspa/Monero sources, while configs/mining_v2.toml declares Qubic as primary.
models/mining_v2/parameters.mem	Q8.8 thresholds — an arithmetic ramp (0x0120…0x0198, step 8), inconsistent with the uniform 1.0 (0x0100) in snn_model.json
models/mining_v2/parameters_weights.mem	Q8.8 synaptic weights
models/mining_v2/parameters_decay.mem	Q8.8 membrane decay rates

These are not reproducible from the data in this repository and show signatures consistent with untrained placeholders. They are under audit. Do not treat them as trained parameters.

---

🔍 Schemas

gpu_telemetry — neuromorphic_data.jsonl

{
  "row_index": 500000,
  "vddcr_gfx_v": 0.7,
  "vram_temp_c": 42.0,
  "gpu_temp_c": 34.0,
  "power_w": 13.104,
  "gpu_clock_mhz": 262.0,
  "mem_clock_mhz": 810.0,
  "clock_mhz": 262.0,
  "fan_speed_pct": 0.0,
  "mem_util_pct": 6.0,
  "qubic_tick_trace": 0.670315,
  "qubic_tick_rate": 0.0,
  "qubic_epoch_progress": 0.0
}

This source carries no timestamp. The collector never emitted one, so records are ordered (row_index) but not time-located. A synthetic clock is deliberately not supplied.

mining — node_sync_harvest.jsonl

{
  "timestamp": "2026-03-19T12:34:56.789000",
  "blockchain": null,
  "block_height": null,
  "chain_epoch": null,
  "hashrate_mh": 1.0,
  "power_w": 125.95439,
  "gpu_temp_c": 67.40456,
  "reward_hint": 0.31488597,
  "qubic_tick_trace": 1.0,
  "qubic_tick_rate": 0.0,
  "qubic_epoch_progress": 0.31488597
}

Chain attribution comes from the source's timestamp field, which takes four forms:

Rows	Source form	Result
114,238	2026-03-19 11:55:13.132	real timestamp, blockchain: null
5,001	dynex:919876	blockchain: "dynex", block_height: 919876, timestamp: null
1,083	qubic:204:46075040	blockchain: "qubic", chain_epoch: 204, block_height: 46075040, timestamp: null
12	2026-03-19 17:00:05.551-05:00	excluded — see below

114,238 rows carry no chain label — that information does not exist anywhere in the source. They are null, never "".

The 12 excluded rows are not capture data. They sit at the end of the source file timestamped ~21 hours before the record they follow, are the only rows in the file carrying a UTC offset, and hold two distinct (power_w, gpu_temp_c) pairs between them: (0.0, 80.0) and (400.0, 40.0). They are placeholders written by a different process, and training on them teaches an idle state that never occurred. The pipeline quarantines them and reports the count; a backward jump in time that it fails to quarantine fails the build instead.

hft — ghost_market_log.jsonl

{
  "timestamp": "2026-03-11T18:22:37.433458521+00:00",
  "step": 1,
  "action": "observe",
  "asset": "PORTFOLIO",
  "price_usd": 70000.0,
  "quantity": 86.0,
  "trade_value_usdt": 0.02700625,
  "realized_pnl_usdt": 0.0,
  "balance_usdt": 500.0,
  "cumulative_pnl": 0.0,
  "portfolio_value": 500.0,
  "ch2_mode": "NVDA",
  "reason": "Warm-up"
}

Actions: buy 14,603 / sell 14,580 / observe 2,390. Assets: PEPE, BTC, RENDER, SOL, NEAR, DNX, ASI, plus PORTFOLIO rows. Simulated trading — not live capital.

qubic_ticks — qubic_ticks_snn.jsonl

{
  "timestamp": "2026-03-20T08:55:24+00:00",
  "tick": 46538099,
  "tick_rate": 0.4333,
  "qubic_tick_trace": 0.0,
  "hashrate_mh_derived": 1.0,
  "power_w_derived": 300.0,
  "gpu_temp_c_derived": 60.0,
  "reward_hint_derived": 0.0
}

The _derived columns are not measurements. They are a fixed function of tick_rate, kept for continuity with earlier consumers. The independent signals here are tick_rate and qubic_tick_trace. An earlier revision published these as hashrate_mh / power_w / gpu_temp_c without qualification; they were never GPU readings.

---

📈 Provenance

rmems/Theseus-Quarry           Rust collectors → raw JSONL
        ↓
rmems/spikenaut-telemetry-etl  ingest → validate → clean → publish
        ↓
rmems/Spikenaut-SNN-Telemetry  this dataset
        ↓
rmems/Spikenaut-SNN            model training

Every file here is generated. Do not hand-edit them; report data issues against the ETL repository, which gates each 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.

---

📝 Changelog

2026-08-04 — 12 placeholder rows removed from mining

node_sync_harvest.jsonl ended with 12 rows that were not capture data: appended out of order (~21h before the record preceding them), the only rows carrying a UTC offset, and holding just two distinct (power_w, gpu_temp_c) pairs.

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 had 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.

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 short out-of-order runs 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 values — inspecting values is what made the original heuristic misfire.

mining 120,334 → 120,322; dataset total 993,310 → 993,298. Chain attribution is unchanged (dynex 5,001 / qubic 1,083); the 12 rows were among the unattributed, now 114,238. Samples are also now emitted in chronological order — a shuffled draw is not a sample of a capture. Other configs are unaffected.

2026-08-03 — data rebuilt from recovered originals

Two files were entirely content-free in prior revisions and have been rebuilt:

File	Before	After
neuromorphic_data.jsonl	813,973 rows, all the identical empty record {"telemetry":{}}	813,973 rows, 12 sensor columns
node_sync_harvest.jsonl	120,314 rows, every numeric 0.0, blockchain ""	120,334 rows, real values and chain attribution

Cause. The previous cleaning script read JSON into a Symbol-keyed dictionary but filtered and looked up fields with String keys. Julia's in and get(d, k, default) fail silently, so every field filter matched nothing and every field lookup returned its default. Nothing asserted the output was non-degenerate, so it reported success.

The same script also overwrote 114,250 real timestamps with a generated base + 10s × index sequence, because its ISO-8601 check tested for a literal T while the real timestamps used a space separator. The previously advertised 2026-03-20 → 2026-04-02 mining window was that fabricated sequence, not an observation period; the real window is 2026-03-19 11:55 → 2026-03-20 14:05. Its "trim synthetic tail" heuristic then fired on its own corrupted output and removed 20 valid rows, which are now restored.

Also in this release:

• Chain attribution recovered from the source timestamp: dynex 5,001, qubic 1,083. A prior card advertised a six-coin breakdown (Kaspa ~45,000, Monero ~38,000, …) that the data never supported; 114,250 rows carry no chain label at all.
• Qubic hashrate_mh / power_w / gpu_temp_c renamed with a _derived suffix. They are a function of tick_rate (16 distinct values across 27,430 rows), not hardware readings.
• Samples regenerated as seeded random draws. They were byte-exact prefixes, which made them unrepresentative and double-counted rows against their parent files.
• 17 dead columns dropped from GPU telemetry and 13 from mining telemetry (constant or all-null in the source, including *_z_score, ocean_intel, solver_*, and the kaspa_* / monero_* drift fields).
• Viewer configs split one-per-file; no config mixes schemas or lists a subset of another file.
• Model artifacts and the previously published 95.2% accuracy figure flagged as unverified.

If you pulled a revision before 2026-08-03, re-download. ghost_market_log.jsonl and qubic_ticks.jsonl were unaffected throughout.

---

📜 Citation

@dataset{spikenaut_snn_v2_telemetry,
  author={Montoya Cardenas, Raul},
  title={Spikenaut SNN v2 Telemetry Dataset},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/rmems/Spikenaut-SNN-Telemetry}
}

---

⚖️ License

MIT OR Apache-2.0 — see LICENSE. Dual-licensed: use whichever fits your project.

---

🙏 Acknowledgments

• Kaspa, Monero, Qubic, Quai, Dynex, Verus communities for open-source node implementations
• E-prop authors (Bellec et al., 2020) for the learning algorithm
• STDP pioneers (Bi & Poo, 1998) for the biological foundation

---

Built by Raul Montoya Cardenas — WGU AI Engineering