Datasets:

---

sagitarim

/

torchlean-benchmark-assets

Like 1

Adds the nine recurrent (LSTM/GRU) fixtures read by the recurrent-v1 adapter of the exact-runtime competition (conjectures-io/conjectures-optimisation-float-kernels-new-competition PR #6).

• New: models/recurrent/** (9 weights + manifests), samples/recurrent-v1/** (7 input pools), ATTRIBUTION.md.
• Changed: manifest.json, regenerated by scripts/index-benchmark-assets.py (21 -> 30 fixtures; the 21 existing entries are byte-identical), and README.md (card updated for the recurrent fixtures, licence metadata added).
• No existing weights, sample pools or fixture manifests are touched or removed.

The MNIST-rows and Appliance recurrent pools are byte-identical to the existing image-zoo-v1/mnist/bc9f8f501f39 and appliance-mixer-v1/.../65edd3bb6d77 pools, so they cost no extra storage.

Every fixture was scored on its full pool by the bit-exact Lean spec (realEval); each spec score matches PyTorch's float32 result, with 100% argmax agreement and outputs within 6e-5.