Skip to content

Dataset cards (lab)

Task: W6.T04 (MNIST / CIFAR / synth audio) + W6.T05 (seq reverse). Datasets are not committed under data/ (gitignored).

MNIST

Use STE MLP pedagogy (bnn train)
Loader bnn MNIST helper (no torchvision required)
Split Standard train/test
License Yann LeCun MNIST terms (public research use)
Card honesty Accuracy floors in tests/golden_floors.json — not SOTA chase

CIFAR-10

Use Bi-Real / image lane (bnn train-image, train-cifar)
Loader HF datasets or local cache under data/ (untracked)
Card honesty Proxy schedule; ImageNet SOTA is a non-goal

Synthetic audio tones

Use Audio lane pedagogy (bnn train-audio)
Generator bnn.audio synthetic tones
Non-goal Production ASR / Whisper replacement

Seq reverse task

W6.T05. Not a downloaded corpus — generated in-process. Do not commit sequences under data/.

Use Encoder–decoder STE pedagogy (bnn train-seq2seq)
Generator On-the-fly reverse sequences (bnn.seq.make_reverse_batch)
Split Synthetic batches each run; no train/test files on disk
License N/A (generated; not a third-party dataset)
Card honesty Smoke: eval token acc = 1.0 at 80 STE steps (results/seq2seq_encoder_decoder.json). Not a real NLP / translation benchmark. Attention and LayerNorm stay FP; FFN is binary or ternary.
Docs 36_ENCODER_DECODER_AND_NEXT.md, tutorial 06, api/seq.md