EntropyCoder/adaptercast-lora-trajectories

AdapterCast LoRA Trajectory Corpus

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).

Layout (per run)

• trajectory.dat.zst — zstd fp16 memmap, shape (n_states, n_params), flat named_parameters() order (NiNo SGDDataset-compatible); factor layout in meta.json (param_specs). Snapshot cadence: see config.json (snapshot_every).
• metrics.npz — per-snapshot σ spectra (T, modules, r), balancedness ‖BᵀB−AAᵀ‖_F, power sums, train/eval loss, probe log-probs; meta_json echoes pinned HF dataset revisions + library versions for exact reproduction.
• config.json / meta.json — full run config; snapshot bookkeeping.

Families

c1_* E1 meta-training corpus (Qwen3-0.6B-Base, r∈{4,8,16}, 12 tasks + 9 canonical mixtures, 600 steps) · e2_* optimizer/long-horizon dynamics arms · t1/2a/2b/2bx_* the E0 measurement program (incl. 1.7B/4B scale rows and gauge-twin experiments).

Held-out protocol

manifest.json echoes the frozen train/held-out splits (tasks, ranks {32,64}, scale) — pre-registered BEFORE meta-training (git tags e1-freeze, e2-freeze).

Note: smoke_* pipeline-validation runs are inventoried in manifest.json for completeness but are not shipped in this dataset.

Reproduction

Code: the adaptercast repository (deterministic pipeline — replicate runs are bit-identical on one host; dataset revisions pinned in each run's provenance).

License: Apache-2.0. Base model: Qwen3-0.6B/1.7B/4B-Base (Apache-2.0). Source datasets retain their upstream licenses (see provenance).