This directory describes the public API, model-family contracts, release
evidence, and contributor workflows for FastPLMs 1.0. The model manifest at
src/fastplms/models.toml is authoritative if prose or generated data differs.
| Goal | Read |
|---|---|
| Understand repository boundaries and loading flow | Architecture |
| Find a supported checkpoint or AutoClass | Models and the generated support matrix |
| Trace every capability to docs, examples, and tests | Capability-to-evidence manifest |
| Embed sequences or FASTA datasets | Embedding API |
| Select SDPA, Flex Attention, or a pinned FlashAttention kernel | Attention backends |
| Build and validate an offline Hub artifact | Artifacts |
| Run parity, structure, or release tests | Testing |
| Measure throughput or memory | Benchmarking |
- ESMFold2: folding, learned representations, the distinct full and Fast MSA contracts, BF16, and experimental FP8.
- Test-time training: opt-in low-rank adaptation and its evidence boundary.
- Binder design: differentiable ESMFold2 and ESM++ research example.
- Fine-tuning: Trainer, PEFT, data splits, and reproducibility.
- Vector benchmark embeddings: reusable embedding artifacts for Protify evaluation.
- Runnable examples: local-only, offline-safe commands for embeddings, attention, generation, RAG, TTT, and structure preparation.
- Contributing: adding models, tests, docs, and examples.
- Licensing: project, source, and checkpoint terms.
- Migration to 1.0: intentional API and repository-layout changes.
Generated files carry a marker stating that they come from
src/fastplms/models.toml. Edit the typed manifest or renderer, then run:
PYTHONPATH=src python -m tools.artifacts.generate_docs
PYTHONPATH=src python -m tools.artifacts.generate_docs --checkDocumentation examples separate model output from experimental evidence. Structure confidence, language-model likelihood, and generated sequences are prioritization signals. Use an independent experiment to establish a biological claim.