Dense LoRA fine-tuning trajectories (adapter weight snapshots + gauge-invariant
spectral features) for research on fine-tuning dynamics and trajectory forecasting.
Companion dataset to the AdapterCast paper (theory: GL(r) gauge structure of LoRA;
measurements: balancedness decay law, Adam pump, regime map; machine: the
NeuralGraphLoRA forecaster).
trajectory.dat.zst — zstd fp16 memmap, shape (n_states, n_params)… See the full description on the dataset page:
https://huggingface.co/datasets/EntropyCoder/adaptercast-lora-trajectories.
