How zoenguyenramirez/arc-prize-2025 works
Project type and purpose
• This is a research‐grade, full-stack training pipeline for a “Pseudo-Reinforcement Learning” framework applied to ARC tasks.
• It orchestrates multi‐iteration model training, self-supervised trajectory generation, data management, and monitoring to improve abstract reasoning models.
File structure
• Top-level folders:
• src/ → Core Python modules (data loaders, model, training orchestrator, inference, utils)
• scripts/ → Bash and Python scripts driving end-to-end pipelines (training, evaluation, data prep, monitoring)
• cloud/ → Helpers for local ↔ remote environment setup, SCP/SSH, TensorBoard on a server
• input_data/ → Raw JSON datasets (ARC-2024, ARC-2025, B-ARC, Re-ARC)
• intermediate_data/ → Prepared .pth datasets for fast loading
• runs/ → Generated checkpoints, TensorBoard logs, trajectories per iteration
• notebooks/ → Exploratory analysis, debugging, visualization
• tests/ → Unit and integration tests validating data prep, training, inference, scripts
• Key files for architecture: orchestrate_training.py, train.py, generate_rollout.py, checkpoint_handler.py, submission.py, hyperparameter_tuning.py.
README for additional insights
• Confirms a 24‐iteration RL-style loop: each iteration trains on combined “original + previously generated trajectories,” then emits new trajectories.
• Emphasizes TensorBoard for monitoring, a “scratch space” of ATTEMPT_START/END tokens, and a detailed run directory layout under runs/.
1. Data Sources
• input_data (ARC-2024/5, B-ARC, Re-ARC JSON)
2. Data Preparation
• prepare_data.py → intermediate_data/*.pth
3. Training Orchestrator
• orchestrate_training.py
• train.py, rotational_train.py, meta_adaptive_train.py
• hyperparameter_tuning.py
4. Model & Checkpoint Management
• model.py, checkpoint_handler.py
• runs/iterN/…/*.pt (checkpoints, optimizer & scheduler states)
5. Trajectory Generator
• generate_rollout.py → runs/trajectories/trajectories_iterN.pt
6. Self-Supervised Loop
7. Monitoring & Logging
• TensorBoard (scripts/start_tensorboard.sh) reading runs/
8. Deployment & Evaluation
• submission.py, sample module for inference
• scripts/evaluation_test.sh, compare_inference.sh
9. Cloud/Remote Infrastructure
• cloud/local_scripts & cloud/remote_scripts enabling SSH, SCP, environment setup
10. Testing & CI
• tests/unit & tests/integration
• Raw JSON → Data Preparation → Prepared .pth
• Prepared .pth + trajectories_iterN-1.pt → Training Orchestrator → New Checkpoint (.pt)
• New Checkpoint → Trajectory Generator → trajectories_iterN.pt
• trajectories_iterN.pt → appended to training data input for iteration N+1
• Training & Generation produce logs → runs/ → TensorBoard
• scripts/ wrap calls to orchestrator, training, evaluation
• cloud scripts interact with runs/ remotely (SSH/SCP)
• tests/ validate scripts & src components
c. Architectural patterns & principles:
• Pipeline + feedback loop (iterative refinement)
• Clear separation of concerns: Data Prep / Training / Inference / Logging / Deployment / Remote Ops
• Modular code in src/, orchestration in scripts/
• Self-supervised “scratch space” via special tokens
d. Technologies & frameworks:
• Python, PyTorch (.pt model files)
• JSON for datasets, .pth for serialized datasets
• Bash shell scripts
• TensorBoard for visualization
• SSH/SCP for cloud orchestration