Datasets:
Telemetry validation pipeline: catch bad data before SNN training
#2
by rmems - opened
Problem
Telemetry data sometimes had bugs that went unnoticed during training. Corrupted inputs (stuck values, NaN, zero deltas, out-of-range) can push all SNN weights in the same direction, producing the monotonic weight artifact seen in the model.
Data Sources at Risk
- Hardware telemetry (GPU NVML): Sensor reads can return stale or NaN values
- HFT market data: Exchange API disconnects produce gaps or zero-fill
- Crypto sync nodes: Node crashes produce incomplete epoch data
- FPGA supervisor: Online learning on hardware may have fed corrupted readings
Proposed Solution
- Pre-ingestion validator: Check every telemetry record for NaN, Inf, stuck values, zero deltas, out-of-range before it reaches the SNN
- Anomaly detection: Flag sudden jumps or flatlines in hashrate, power, temperature
- Training logging: Save telemetry snapshots alongside weight updates so bad data can be traced
- Retrospective analysis: Audit existing training logs for data quality issues
Expected Impact
Clean telemetry in means clean weights out. This is the most likely root cause of the monotonic weight pattern.
Mimo Code agent: MiMo-V2.5-Pro