sagitarim/torchlean-benchmark-assets

TorchLean benchmark assets

The real-model fixtures used by the TorchLean exact-runtime benchmark. Each fixture is a trained model's weights plus a pool of held-out inputs. The benchmark downloads a pinned commit of this repository into data/benchmark/assets/ of the competition checkout. That directory is Git-ignored there and is the root of this dataset.

Layout

• models/image-zoo/<family>/<dataset>/: trained image-model weights (weights.bin) and the fixture's manifest.json.
• models/uci-har/<family>/ and models/appliance-energy/mixer/: weights and manifests for the other non-recurrent fixtures.
• models/recurrent/<dataset>/<model>/: recurrent (LSTM/GRU) weights and manifests, read by the recurrent-v1 adapter.
• samples/<adapter>/<dataset>/<content-id>/samples.bin: the complete held-out input pool shared by every fixture that uses that exact sample set. Labels are omitted. <content-id> is the first 12 hex digits of the file's SHA-256.
• manifest.json: generated catalog of all fixtures with relative paths, checksums, sizes and dimensions.
• ATTRIBUTION.md: licence and attribution for every source dataset.

Binary tensor data is little-endian float32. Fixture manifests record the adapter and workload identity, tensor and sample dimensions, provenance and SHA-256 hashes. The benchmark configuration pins every fixture's weight and sample hashes and verifies them before measuring.

Fixtures

Adapter	Fixtures	Input pools
image-zoo-v1	MLP, normalised MLP, ResNet, MobileNet, Mixer on sklearn Digits, MNIST and Fashion-MNIST (15)	360 Digits, 10,000 MNIST test, 10,000 Fashion-MNIST test images
uci-har-v1	MLP, normalised MLP, ResNet, MobileNet, Mixer (5)	2,947 UCI-HAR test windows
appliance-mixer-v1	Mixer regression (1)	2,932 chronological Appliance Energy test windows
recurrent-v1	9 LSTM/GRU models, listed below	full test split of each dataset

Recurrent fixtures (recurrent-v1)

Each recurrent model is evaluated on the whole sequence, and only the hidden state of the last step feeds a linear head. Each sample is the T × F input sequence, step by step. The first tensor of weights.bin is a header that describes the model, stored as unsigned 32-bit integers rather than floats: [1, T, F, outputs, L, kind_1, H_1, …] (kind 0 = LSTM, 1 = GRU). Then come each layer's w_ih, w_hh, b_ih, b_hh in PyTorch's layouts and gate orders, then the head's weight and bias. Every fixture was scored on its full pool by the bit-exact Lean specification (realEval). Each score below matches PyTorch's float32 result (accuracy exactly, RMSE to 9 digits). The two agree on the predicted class for every sample, and no output differs from PyTorch's by more than 6e-5.

Workload	Layers (hidden 32)	Steps × features → outputs	Pool	Spec result
recurrent-uci-har-lstm	LSTM	64 × 9 → 6	2,947	89.51% accuracy
recurrent-uci-har-gru	GRU	64 × 9 → 6	2,947	89.82% accuracy
recurrent-appliance-gru	GRU	24 × 8 → 6 (regression)	2,932	RMSE 0.7317
recurrent-mnist-rows-lstm	LSTM	28 × 28 → 10	10,000	98.60% accuracy
recurrent-spoken-arabic-digit-lstm	LSTM	48 × 13 MFCC → 10	2,200	95.86% accuracy
recurrent-tiny-shakespeare-lstm	LSTM	64 × 65 one-hot → 65	1,742	39.90% next-character accuracy
recurrent-mitbih-gru	GRU	94 × 1 → 5	21,829	96.47% accuracy
recurrent-speech-commands-gru	GRU	49 × 40 MFCC → 12	4,890	88.92% accuracy
recurrent-speech-commands-lstm-gru	LSTM → GRU	49 × 40 MFCC → 12	4,890	89.16% accuracy

Each recurrent manifest lists the architecture, preprocessing, splits, licence, attribution and the training and spec metrics. The training and export code is in training/recurrent/ of the competition repository.

Verify the catalog

After adding or regenerating fixtures, run this from the competition checkout:

python scripts/index-benchmark-assets.py

It checksum-verifies every referenced file and rewrites manifest.json. scripts/split-real-fixtures.py rebuilds the non-recurrent fixtures, and scripts/build-recurrent-fixtures.py rebuilds the recurrent ones.

Use in the benchmark

scripts/download-benchmark-assets.py --revision COMMIT

This needs HF_TOKEN_TORCHLEAN_BENCHMARK in the environment or .env. Then generate the configuration with scripts/exact-config.py and any of --real-image-zoo, --real-uci-har, --real-appliance-mixer or --real-recurrent. Override the location with --assets-dir PATH or TORCHLEAN_BENCHMARK_ASSETS=PATH; it must be inside the checkout so the worker can stage the assets into its sandbox. Pin an immutable Hub commit for deployment; never deploy from main.

Licences

The input pools are preprocessed held-out data from public datasets under CC BY 4.0, ODC-By 1.0, CC BY-SA 3.0 (MNIST) and MIT (Fashion-MNIST, char-rnn) terms. See ATTRIBUTION.md for each dataset's licence and citation. The repository holds only runtime weights, sampled inputs, manifests and documentation: no training intermediates, full raw datasets, logs or compiled outputs.