Experimental model training

Status: experimental. The distill/ directory contains teacher-sample generation, chat-data preparation, and QLoRA adapter training. These are separate from the serving backend.

What exists

The experiment documents 16 teacher examples, split into 15 training examples and one validation example. It reports a short-context training trial and saved adapters. Treat those results as recorded experiment evidence, not a guarantee that a rerun succeeds or that planning quality improved.

Before attempting a run

• Use a dedicated training environment; do not upgrade the live backend’s Python packages.
• Verify the exact base-model identifier, architecture, tokenizer, and adapter compatibility. The scripts and README make model-identity claims that need independent verification.
• Check current GPU memory requirements for your chosen sequence length and batch configuration.
• Inspect training examples for confidential data and licensing restrictions.
• Read distill/README.md and the scripts at the revision you intend to run.

Evaluate a candidate

Compare baseline and trained models on held-out tasks that were not used for teacher generation or tuning. Measure schema validity, completeness, invented paths/commands, acceptance-criteria quality, and execution safety. Record latency and resource use separately from answer quality.

A single validation example and a lower training loss cannot establish generalization. Require a reproducible evaluation before replacing a serving model. Merge/export and Ollama registration are further steps, not implied by an adapter file existing.

Next: Planner handoff contract and project roadmap.