masterpieceexternal/gpt-oss-20b-moe-expert-power-traces-320k-ds16k

GPT-OSS-20B MoE Expert Power Traces (Downsampled to 16k)

Downsampled variant of the 320k expert-trace capture set.

Source

Raw source dataset (same captures):

• 32 experts (expert_00..expert_31)
• 10,000 traces per expert
• 320,000 total traces
• raw trace length ~195k samples per trace

Downsampling method

Each raw trace was resampled to exactly 16384 samples using linear interpolation (np.interp) matching the trainer resampling step.

No baseline normalization and no derivative feature are pre-baked; those are applied at train time.

Format

For each class folder expert_XX:

• traces.npy: shape (10000, 16384), dtype float32
• trial_ids.npy: shape (10000,), dtype int32 (0..9999)
• meta.json: class-level metadata

Top-level files:

• capture_meta.json
• downsample_summary.json
• verify_summary.json (integrity/split check output)
• scripts/downsample_traces_to_16k.py
• scripts/reproduce_cnn52_from_ds16k.py
• scripts/verify_ds16k_dataset.py
• scripts/make_smoke_subset.py
• scripts/train_expert_classifier_multiclass.py

Quick start

1) Download

pip install -U "huggingface_hub[cli]"
hf download masterpieceexternal/gpt-oss-20b-moe-expert-power-traces-320k-ds16k \
  --repo-type dataset \
  --local-dir ./gpt-oss-ds16k
cd ./gpt-oss-ds16k

2) Verify dataset integrity and grouped split

python scripts/verify_ds16k_dataset.py \
  --dataset-root . \
  --expect-classes 32 \
  --expect-traces-per-class 10000 \
  --expect-len 16384 \
  --val-size 10000 \
  --seed 7

Expected split preview:

• train_size = 310016
• val_size_actual = 9984
• val_group_count = 312

3) Smoke test that training script runs

python scripts/make_smoke_subset.py \
  --dataset-root . \
  --out-root /tmp/ds16k_smoke_subset \
  --per-class 64

python scripts/reproduce_cnn52_from_ds16k.py \
  --dataset-root /tmp/ds16k_smoke_subset \
  --run-dir /tmp/repro_cnn52_smoke \
  --epochs 1 \
  --batch-size 64 \
  --lr 1e-3 \
  --warmup-epochs 0 \
  --seed 7 \
  --feature-len 16384 \
  --baseline-samples 2000 \
  --val-size 512 \
  --num-workers 4 \
  --expect-min-val-acc 0.0

Full reproduction (~52% regime)

python scripts/reproduce_cnn52_from_ds16k.py \
  --dataset-root . \
  --run-dir ./repro_cnn52 \
  --epochs 20 \
  --batch-size 32 \
  --lr 1e-3 \
  --warmup-epochs 2 \
  --seed 7 \
  --feature-len 16384 \
  --baseline-samples 2000 \
  --val-size 10000 \
  --num-workers 8 \
  --expect-min-val-acc 0.50

Hyperparameters mirror the reference run:

• preprocess: baseline z-score (first 2000 samples) + dx concat ([x, dx] => input dim 32768)
• split: grouped by trial_ids, seed 7, target val size 10,000 (actual 9,984)
• model: CNN with BatchNorm (track_running_stats=True), dropout 0.2
• optimizer/schedule: Adam (lr=1e-3, weight_decay=0), cosine decay, warmup 2 epochs
• training: 20 epochs, batch size 32

Notes

• Controlled forced-single-expert harness capture, not a full unmodified forward pass.
• Validation accuracy has natural run-to-run variance from stochastic training/CUDA behavior.