Fine-tuning further

Every new pass must name an immutable parent. Do not replace a stable adapter in place.

1. Add licensed records with record-level provenance.
2. Bump the dataset version and rebuild its manifest.
3. Add fresh held-out evaluation cases without copying training prompts.
4. Copy the training config to a new experiment ID.
5. Set parent_checkpoint to the stable adapter and base revision.
6. Evaluate the parent before training.
7. Train, evaluate, blind, gate, and record the outcome.

Base a07cc9a…
└── EXP-…-001 → rejected or TideLM v0.1.0
    └── EXP-…-002 → rejected or TideLM v0.2.0

The current runner supports the audited base as the first parent. Loading an existing parent adapter for a subsequent pass is an intentionally tracked next feature; do not imply lineage support until its implementation and tests land.

Data contributions

Use data/schema.json. Compatible licenses are allowlisted, but allowlisting is not a substitute for checking each source's terms. Keep independent test data out of data/raw, document synthetic generators, and never label AI-only review as human expert review.

Choosing LoRA scope

Expanding to more layers increases capacity and activation memory. Record rank, alpha, dropout, exact module names, trainable count, precision, optimizer, steps, and hardware. Compare against the parent under the same generation settings. A lower training loss alone never earns promotion.