PDLI Experiment — Multi-Project ML Research Workspace

A shared research environment for multiple machine learning projects, with a common paper library, technique implementations, and tooling.

Projects

Project	Directory	Description
PDLI	projects/pdli/	Positional Dispositional Layer Insertion — training alignment layers in GPT-2
LTP	projects/ltp/	Latent Thought Pipeline — masked thought rendering
Music-Sim	projects/music-sim/	Music similarity via neural embeddings and PMI
Distill	projects/distill/	Multi-teacher latent distillation

Quick Start

# Activate venv
source .venv/bin/activate    # Linux
.\.venv\Scripts\activate     # Windows

# Run PDLI experiments
python run.py 01             # Single experiment
python run.py 01 02 03       # Multiple
python run.py all            # All four standard notebooks

Shared Library

The shared/ directory provides cross-project resources:

• shared/docs/upstream/papers/ — Paper summaries and PDFs
• shared/resources/techniques/ — Paper technique implementations
• shared/references/papers.md — Master paper index

Structure

.
├── projects/          # Self-contained research projects
│   ├── pdli/          # Each with its own notebooks, components, runs
│   ├── ltp/
│   ├── music-sim/
│   └── distill/
├── shared/            # Cross-project library (docs, techniques, references)
├── tools/             # Utility scripts
├── artifacts/         # Workspace meta (logs, decisions, plans)
├── mlruns/            # Shared MLflow database (SQLite)
├── .opencode/         # Agent configuration
└── lesson_plans/      # Agentic tooling (do not modify)

Environment

• Python 3.12, venv at .venv/
• PyTorch 2.6.0+cu124 (CUDA 12.4)
• MLflow with SQLite backend at mlruns/mlflow.db