BDH development

[ganeshflexiana]

No description provided.

[ganeshflexiana]

Baby Dragon Hatchling (BDH) Training Pipeline

Adds an end-to-end pipeline for training, evaluating, and demonstrating memory on a small character-level LM.

What's new:

• train.py – configurable training with 10%-interval checkpointing
• compare_checkpoints.py – checkpoint evaluation report and generation log
• memory.py – fast (context-based) and model (weight-based) memory demos
• main.py – one-command runner for the full pipeline
• Docker images for CPU and GPU (CUDA)

To run locally:

pip install numpy torch requests psutil pandas matplotlib
python main.py --max_iters 1000

⚠️ Local Memory Note: Running locally requires more memory than the Docker setup. Make sure to configure your system memory settings before running — closing other applications and starting with a smaller --max_iters (e.g. 100) is recommended to test memory usage first. Use Docker if your machine has limited RAM.

Outputs land in outputs/training/, outputs/evaluation/, and outputs/memory/. See README.md for full details on interpreting the loss curve and checkpoint samples.