Repository navigation
🌱 Bonsai v0.1.0 — Initial Public Release
The art of cultivating massive foundation models into pocket-sized intelligences on consumer silicon.
uvx bonsai prune ./data✨ Key Capabilities
- Hardware Auto-Probe: Probes GPU VRAM, Apple Unified Memory, or CPU RAM and auto-downshifts base models to guarantee Zero-OOM.
- Zero-Leakage Scenario Partitioner: De-duplicates input scenarios before augmentation so the 15% blind holdout exam is strictly unseen (Jaccard similarity = 0.0).
- Masked-Loss Pocket LoRA: Computes loss strictly on assistant completions (
labels[:prompt_len] = -100), reducing loss by -51.5% in 34 seconds on a consumer laptop RTX 2050 (4GB VRAM). - The Proof Machine: Automatically evaluates the untrained base model side-by-side against the cultivated model on unseen test cases.
- Instant Export: Emits ready-to-run Ollama Modelfiles, standalone Python runners, and local Web UI (
bonsai ui).
🧪 Verified Benchmarks (RTX 2050 - 4GB VRAM)
- Base Model:
Qwen2.5-Coder-0.5B-Instruct(495M parameters) - Peak VRAM Allocated: 2,045 MB (Zero-OOM Protected)
- Distillation Duration: 34.05 seconds (3 epochs)
- Initial Loss: 1.5905 ➔ Final Loss: 0.7713 (-51.5% Drop)
- Automated Test Suite: 11/11 Passing via pytest