CoolFace
Datasetpublic

masterpieceexternal/gpt-oss-20b-lmsys-layer2-expert-traces-600k-10msps

GPT-OSS-20B LMSYS Layer-2 Expert Power Traces — 600k tokens at 10 MSPS This dataset contains real power/current traces captured from an NVIDIA H100 PCIe system while running GPT-OSS-20B decode on LMSYS prompts. The capture target is layer 2 MoE expert execution only; earlier/later transformer work is executed off-trace to avoid wasting acquisition time. Capture summary Model: openai/gpt-oss-20b Target layer: 2 Captured region: target_layer_experts_only Prompts:… See the full description on the dataset page: https://huggingface.co/datasets/masterpieceexternal/gpt-oss-20b-lmsys-layer2-expert-traces-600k-10msps.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
0likes118downloads
Dataset Card

GPT-OSS-20B LMSYS Layer-2 Expert Power Traces — 600k tokens at 10 MSPS

This dataset contains real power/current traces captured from an NVIDIA H100 PCIe system while running GPT-OSS-20B decode on LMSYS prompts. The capture target is layer 2 MoE expert execution only; earlier/later transformer work is executed off-trace to avoid wasting acquisition time.

Capture summary

  • —Model: openai/gpt-oss-20b
  • —Target layer: 2
  • —Captured region: target_layer_experts_only
  • —Prompts: LMSYS Chat, English-filtered prompt plan
  • —Accepted decode tokens: 600,000
  • —Prompts touched: 9,375
  • —Experts per token: 4
  • —Total expert windows: 2,400,000
  • —Unique experts observed: 32 / 32
  • —Trace sample rate: 10 MSPS
  • —Samples per capture: 100,000
  • —Capture duration: 10 ms
  • —Trace dtype: float16
  • —Trace shape: (100000,)
  • —Approximate trace payload size: 120.08 GB

Layout

text
traces/shard_XXXXXX/step_YYYYYY.npy      # float16 power trace, shape (100000,)
records/shard_XXXXXX/step_YYYYYY.json    # per-token metadata and expert window bounds
timelines/shard_XXXXXX/...               # periodic timing/timeline artifacts
prompt_plan.jsonl                        # deterministic prompt plan used by the run
capture_meta.json                        # capture summary and expert counts
run_state.json                           # final committed run state
layer_windows.rebuilt.jsonl              # flattened expert-window index from audit

Each record contains:

  • —global_step
  • —prompt_plan_index
  • —token_in_prompt
  • —hit_experts
  • —trace_file
  • —trace_stats
  • —layer_windows, with expert_idx, sample_start, sample_end, and millisecond boundaries
  • —scope configuration including adc_freq_effective, adc_decimate, and adc_samples

Health/audit status

Final audit status:

  • —committed records: 600,000
  • —missing trace files: 0
  • —trace dtypes: float16: 600000
  • —trace shapes: (100000,): 600000
  • —loaded trace sample count during audit: 1,000
  • —audit warnings: 0
  • —audit errors: 0

A separate visual spot check sampled early, middle, and late captures and confirmed that the annotated expert windows are inside the traces and visually aligned.

Notes

The analog chain is AC-coupled/high-pass and the installed current clamp is outside its specified flat bandwidth for MHz-scale components. Treat absolute current calibration cautiously; the main intended use is comparative classification/analysis of expert-conditioned trace structure.