bnn-lab MNIST MLP `.bnnpack` (canary)

Canary / lab demo — not ImageNet SOTA, not a MNIST leaderboard submission.

Packed hidden BinaryLinear layers from bnn.models.BinaryMLP (hidden=512). FP stem (784→512) and FP head (512→10) are not in this pack (standard BNN practice; encode skips them).

This file vs published MNIST numbers

This Hub artifact is a codec canary (seed-0 build_model("binary_mlp") BinaryLinear blobs). It is not the 3-epoch trained checkpoint (checkpoints/binary_mlp.pt is gitignored). Do not report this file's inference accuracy as the lab MNIST result.

Published STE train floors (CPU, seed 42, 3 epochs) from results/train_results.json and tests/golden_floors.json mnist:

Model	Recorded test acc	Floor
`fp32_mlp`	97.67	`fp32_mlp_min_acc`: 96.0
`binary_mlp`	96.36	`binary_mlp_min_acc`: 95.0
`ternary_mlp`	97.16	`ternary_mlp_min_acc`: 95.0

Gap gate: gap_max_pp_fp_vs_binary: 3.0 pp when FP ≥ fp_for_gap_gate 97.0. Retrain + re-encode: bnn train --model binary_mlp --epochs 3 --seed 42 then python scripts/encode_hf_canaries.py --only mnist-mlp --from-checkpoint checkpoints/binary_mlp.pt.

What 32× means

32× is uint64 pack compression of aligned BinaryLinear weights (in_features % 64 == 0), not a GPU speedup from sign(). Training with STE is simulation.

Load

from huggingface_hub import hf_hub_download
from bnn.codec import decode_file, packed_module_fp_err

path = hf_hub_download(
    "KanakMalpani/bnn-lab-mnist-mlp-canary", filename="model.bnnpack"
)
modules, meta = decode_file(path)
for name, mod in modules.items():
    print(name, mod.in_features, mod.out_features, packed_module_fp_err(mod))

Tutorial: docs/tutorials/08_HF_OPTIMISER.md.