Upload hf_dataset_card.md

origin_hardware_baselines/resident_evil_4/hf_dataset_card.md

@@ -0,0 +1,76 @@
+---
+license: gpl-3.0
+task_categories:
+- time-series-forecasting
+tags:
+- neuromorphic
+- snn
+- liquid-state-machines
+- gaming
+- hardware-telemetry
+- gpu
+pretty_name: Metis SMoE Latent Telemetry (Gaming)
+---
+
+# Metis SMoE Latent Telemetry
+## Neuromorphic Hardware Telemetry from demanding Gaming Workloads
+
+This dataset provides high-fidelity, high-frequency (5ms interval) hardware telemetry data captured from extreme PC gaming workloads. This dataset is optimized to simulate the biological responses of a nervous system to intense stimulus (excitatory input, action potentials/firing rates, and inhibitory responses).
+
+### Context
+The telemetry data was recorded using a custom Rust-based data collector via the NVIDIA Management Library (NVML) on a Fedora 43 Linux system. Workloads represent highly transient rendering applications including:
+- **Resident Evil 4 (Remake)** (with rendering complexities)
+- **Cyberpunk 2077** (Path Tracing, DLSS 4.0)
+
+This system provides the rich, high-frequency time-series data required to train **Spiking Neural Networks (SNNs)** and **Liquid State Machines (LSMs)**.
+
+### Neuromorphic Mapping (SNN Utility)
+This data behaves as "sensorimotor" stimulus for neural networks:
+- **Excitatory Inputs (Stimulus):** High surges in `pcie_rx_kbps` indicate asset floods (e.g., BVH structure updates for Path Tracing), mimicking sensory signals entering the system.
+- **Action Potentials (Firing Rates):** `encoder_util_perc`, `decoder_util_perc`, and overall spatial `power_usage_mw` transients represent internal activity and network firing rates.
+- **Inhibitory Inputs (Refractory/Limits):** Non-zero `throttle_reasons_bitmask` signals and thermal limits act as inhibitory governors, dynamically suppressing system activity.
+- **State/Momentum:** Slow-moving environmental data like temperatures (`cpu_tctl_c`, `temperature_c`) and memory capacity.
+
+### Data Schema
+
+The data is provided natively as Parquet files partitioned into train batches (`system_telemetry_v1_batch_*.parquet`).
+
+| Feature | Type | Description |
+| :--- | :--- | :--- |
+| `timestamp_ms` | `Int64` | UNIX timestamp in milliseconds (5ms interval captures). |
+| `power_usage_mw` | `UInt32` | Total GPU power usage in milliwatts. |
+| `temperature_c` | `Float32` | GPU core temperature in Celsius. |
+| `pcie_rx_kbps` | `UInt32` | Incoming PCIe throughput in Kilobytes per second (Excitatory). |
+| `pcie_tx_kbps` | `UInt32` | Outgoing PCIe throughput in Kilobytes per second. |
+| `encoder_util_perc` | `Float32` | NVIDIA Encoder (NVENC) utilization percentage. |
+| `decoder_util_perc` | `Float32` | NVIDIA Decoder (NVDEC) utilization percentage. |
+| `mangohud_active` | `Boolean` | Whether MangoHud overlay telemetry was active during the snapshot. |
+| `cpu_tctl_c` | `Float32` | Primary CPU package temperature (Tctl). |
+| `cpu_ccd1_c` | `Float32` | Temperature of CPU Core Complex Die 1. |
+| `cpu_ccd2_c` | `Float32` | Temperature of CPU Core Complex Die 2. |
+| `throttle_reasons_bitmask`| `UInt64` | Bitmask defining hardware throttling events (Power, Thermal, Sync) - acts as Inhibitory signals. |
+
+### Usage with Hugging Face `datasets`
+
+You can seamlessly integrate this telemetry into your Neuromorphic modeling workflows using the Hugging Face `datasets` library.
+
+```python
+from datasets import load_dataset
+import pyarrow.parquet as pq
+
+# Load the entire telemetry dataset as a single stream
+dataset = load_dataset("rmems/Metis-SMoE-Latent-Telemetry", split="train")
+
+print(dataset.features)
+print(dataset[0])
+```
+
+### Export to Canonical CSV (For Corinth Canal Replay)
+If you are using the Spikenaut `corinth-canal` framework, you can export a canonical CSV by grabbing a single dataset file:
+
+```bash
+cargo run --bin export_csv data/train/system_telemetry_v1_batch_1.parquet canonical.csv
+```
+
+### License
+This dataset is distributed under the GPL-3.0 License.