ML Workshop is a public engineering and research workbench for transformer, state-space language-model, and scientific-ML systems. It preserves inspectable source, evaluation paths, Slurm job definitions, runtime reports, and explicit claim boundaries across experiments.
| Capability | Public evidence |
|---|---|
| PyTorch model systems | Decoder-only transformer components, hybrid state-space/attention blocks, checkpoint loading, post-training objectives, and evaluation code in transformer_lab/ and state_space_slm_lab/ |
| Cluster execution | Registered Slurm jobs for model evaluation, turbulence, MHD, GPU probes, and managed services in scripts/ and the repository root |
| Evaluation and reliability | Generation, next-token, cloze, sentence-suite, artifact compatibility, blocked-runtime, and trace persistence paths in mlworkshop_intel/model_eval.py |
| Service and interface layer | A bounded FastAPI physics service plus Next.js observability and control surfaces in scripts/serve_physics_ai.py and frontend/ |
| Verification | A documented 49-test Python snapshot, frontend lint/build checks, and real checkpoint/tokenizer/GPU smoke evidence |
| Evidence discipline | Model lineages, data provenance, corrections, negative results, and environment-specific limitations are kept explicit |
Start with:
- Model lineages for the two distinct language-model architectures and checkpoint histories.
- Engineering evidence index for direct paths into code, tests, jobs, and runtime records.
- Codebase map for component ownership.
- Validation and risks for test coverage and unresolved boundaries.
| Lineage | Architecture | Durable checkpoint | Current public role |
|---|---|---|---|
small_frontier_dense |
Conventional decoder-only ResearchTransformer, 18 layers, d_model=768, 12 attention heads, 4 KV heads |
step 499999 | Checkpoint inspection and the public model-evaluation cockpit |
state_space_slm_300m |
Separate hybrid recurrent state-space/attention architecture with graph-memory and MoE experiments | step 114000 | Preserved historical rank-zero lineage with 3,735,552,000 checkpoint-effective tokens |
These are different model architectures, artifact paths, and training campaigns. The step numbers are not competing descriptions of one run. See Model lineages for the exact token-accounting correction and evidence links.
ML Workshop connects three bounded streams:
- transformer and state-space language-model systems;
- turbulence and magnetohydrodynamics scientific ML;
- state-native scientific inquiry.
See RESEARCH_REPOSITORIES.md for current ownership,
provenance, and publication status. Shared machine-readable contracts live in
contracts/.
The implementation repositories created after the July 2026 split remain private while dataset licensing, claims, and reproducibility boundaries are audited. This public umbrella retains historical source, public-safe documentation, and read-only observability surfaces. It is not a remote execution service.
All mutating API routes are disabled unless the server has
MLWORKSHOP_LOCAL_MUTATION_TOKEN and the request provides the same value as a
Bearer token. Public deployments should leave this variable unset. Cluster work
is submitted through governed, registered actions and Slurm, never through
anonymous arbitrary command execution.
Historical notebooks, source, and reports remain available so prior work can be audited. Reports describe specific snapshots; executable truth belongs to package code and registries, while run truth belongs to manifests and live runtime evidence.
The historical state_space_slm_300m launcher started two unsynchronized
ranks. Only rank zero wrote the preserved checkpoint. Its model exposure is
therefore reported using checkpoint-effective tokens, not doubled aggregate
hardware processing. Historical corpus provenance and overlap questions also
remain under audit; no held-out generalization claim is made from that lineage.