Qwen Compression Lab

Version-controlled engineering pipeline for compressing the pinned Qwen 3.6 27B checkpoint into a measured local Hermes runtime for a 16GB M4 Mac mini.

Safety rules

• Original checkpoints are immutable and live outside this Git repository.
• Every source is pinned by repository revision and verified by hash.
• No candidate is called usable until peak RAM, swap, speed, stability, and quality are measured.
• Failed experiments remain in docs/EXPERIMENT_LOG.md.
• Hermes configuration remains separate from model weights.
• Large artifacts, caches, datasets, and generated checkpoints are excluded from Git.

Initial commands

python3 scripts/memory_budget.py --output results/memory-budget.json
python3 -m unittest discover -s tests -v

Layout

• manifests/ — pinned source identities and hashes
• docs/reports/ — architecture, research, memory, benchmark, and final reports
• docs/plans/ — executable implementation plan
• scripts/ — reproducible model and benchmark tools
• tests/ — deterministic pipeline tests
• configs/ — candidate/runtime configuration
• results/ — small structured outputs only
• logs/ — command receipts; no secrets

Large artifacts target /Users/richardholguin/Models/Qwen-Compression-Lab and are never committed.