TASK.md

Task List – Flexible Learning PLM Framework

Grouped by work‑stream; tick ☑︎ when done.

---

0 ‒ Project Bootstrap

• Create Git repo; enable GitHub Actions template.
• Define editable Poetry environment (pyproject.toml).
• Pre‑commit hooks: ruff, black, isort, mdformat.

---

1 ‒ Core Infrastructure

• Data schema → datamodels.py (Pydantic) for Variant, AssayResult, RoundMetadata.
• Data manager with SQLite / DuckDB backend; CRUD, versioning.
• Embedding cache with mmap HDF5; script embed.py (batch, multiprocess).
• Config system via yaml + env overrides (OmegaConf).

---

2 ‒ Learner Heads

• Ridge / MLP head using sklearn / torch.nn.
• Adapter‑LoRA utility to unfreeze ≤ 2 layers (lora_r=4).
• RL policy (A2C): action = {pos, aa}, mask illegal edits.

---

3 ‒ Active‑Learning Loop

• Acquisition strategies module: UCB, EI, Thompson, diversity‑penalized.
• controller.py orchestrating propose → assay → fit cycle.
• Temperature & batch‑size scheduling CLI flags.

---

4 ‒ User Interfaces

• CLI (Typer): plm propose, plm learn, plm benchmark.
• Python API: from plm import Framework; Framework().fit_round(...).
• Dashboard (optional): Streamlit view of embedding UMAP & acquisition scores.

---

5 ‒ Evaluation & Benchmarks

• Port Science 2024 & Nat Commun 2025 datasets (convert to internal schema).
• scripts/benchmark.py to generate metrics table & figures.
• Unit tests (> 80 % coverage) for head training & acquisition math.

---

6 ‒ Performance & Optimization

• FP16 / Int8 quantization toggle (bitsandbytes).
• CPU‑only training path; profiling with torch.profiler.
• Automated runtime budget check (< 2 h on 8‑core Intel i7).

---

7 ‒ Documentation & Examples

• MkDocs site with tutorial notebooks ("0‑shot", "1‑round", RL demo).
• API reference auto‑generated via mkdocstrings.
• Contribution guide & code‑of‑conduct.

---

8 ‒ Release Engineering

• Semantic versioning; generate changelog with towncrier.
• Build & publish wheels to TestPyPI → PyPI.
• Dockerfile (CPU slim) for reproducible runs.

---

9 ‒ Stretch Goals

• Multi‑objective optimization (e.g., stability × activity via Pareto front).
• GPT‑4o‑driven candidate filtering (natural‑language constraints).
• Fine‑tune on metagenomic unlabeled sequences (self‑distillation).

---