Goldens

Reference arrays for tests that check a port of a model against the original, number for number. Each folder belongs to one port.

tsfm/

Recordings of IBM's Granite TinyTimeMixer r3, ibm-granite/granite-timeseries-ttm-r3 at revision 512-48-dec-512-lite-r3 (commit 1cb447a4f769e2c8dcc0d1c78de0775610555b2b), made with IBM's own PyTorch code. A JAX port of the model is tested against them, so that the tests need neither PyTorch nor IBM's package.

File	What it holds
ttm_r3_512_48_lite.npz	inputs, the model's outputs, and two families of gradients
ttm_r3_512_48_lite_activations.npz	every leaf module's output on the same inputs
ttm_r3_512_48_lite_step.npz	one AdamW step from the checkpoint, as the Hugging Face Trainer builds its optimizer
ttm_r3_512_48_lite_trajectory.npz	an open-loop optimizer recording and two 100-step training trajectories
ttm_r3_512_48_lite_fresh_init.safetensors	a freshly initialised model at seed 0
tsp_parity.npz	IBM's TimeSeriesPreprocessor fitted on a small synthetic frame
config.json	the checkpoint's configuration, as published by IBM
*manifest.yaml	for each file above: every array's shape, dtype and SHA-256, the checkpoint's revision, and the versions of the packages that produced it

The inputs are synthetic. The manifests record the versions used: granite-tsfm 0.3.9, torch 2.11.0, transformers 5.15.1, numpy 2.5.2, scikit-learn 1.7.2, pandas 3.0.5.

The files derive from IBM's model, which is published under the Apache License 2.0; they are distributed under the same licence.