drift-sampler artifacts
Recorded training checkpoints for an ongoing research project on one-step
samplers for unnormalized Boltzmann targets p(x) ∝ exp(-E(x)).
68 files, 3.2 GB. This repository is a storage backend, not a released model. On its own it does very little: the weights are only meaningful together with the training/inference code, which is not public.
What this is
Checkpoints from three eras of the project, laid out exactly as the code repo expects them:
| prefix | count | what |
|---|---|---|
checkpoints/blessed/ |
34 | 2026-06, RTX 5090 workstation |
checkpoints/hpc_2026-07/ |
30 | 2026-07-29 → 08-01, NUS Hopper H200 benchmark suite |
checkpoints/aldp/ |
4 | 2026-07-31 alanine-dipeptide pipeline run-through |
They are kept because they are not reproducible: the trainer never set
deterministic=True, and these runs degrade late, so the best weights exist
only in an intermediate checkpoint that re-running does not recover.
What this is NOT
- No code. No model definition, no drift kernel, no training loop.
- No method configuration. The checkpoints carry no
hyper_parametersblock — none of the kernel scales, drift coefficients, or objective weights that define the method are present. Only tensors, optimizer state, and PyTorch-Lightning bookkeeping. - No datasets. These are target-density benchmarks; there is no training data. The energy function is the only access to the target.
- No paper. The method is unpublished.
Consequently a state_dict here tells you tensor shapes and little else.
Usage
Fetched by the (private) code repository's own tooling, which verifies every file against a sha256 index:
python scripts/artifacts.py pull
python scripts/artifacts.py verify
Direct download works too, but the paths are only meaningful inside that repo.
Provenance and contact
Produced on the NUS Hopper cluster (PBS, H200) and an RTX 5090 workstation. Per-file provenance — which run produced which weight, and which reported number it corresponds to — lives in the code repository, not here.
This is research work in progress. If you want to use or cite any of it, please get in touch first.