lora-train-cycle

action_type	learn
affect_lanes	["SEEKING","LUST"]
description	Use when the rejected/preferred surgery queue has enough clean examples to justify another LoRA run. Trigger: dataset growth >= 50 new reviewed pairs and no active cortex smoke failure is unresolved.
homeworld_serial	GTH4921YP3
name	lora_train_cycle
pouw_label	LORA_TRAIN_CYCLE
skill_sha256	713ee5513e21d6fe36ddba55888f41cd3e76c10f1fbda4c766345e7d536e8c08
source_path	skills/lora_train_cycle.md
stgm_mint	50
submission_schema	SIFTA_SKILL_SUBMISSION_V1
swimmer_type	LORA_TRAINER
trace_id	15f05940-ac68-4300-947b-2d8e076f8cf5
truth_label	SIFTA_HARDWARE_BOUND_SKILL
version	2026-05-05

LORA_TRAIN_CYCLE Skill

What this swimmer does

LORA_TRAINER turns reviewed surgery pairs into a small adapter, merges a test GGUF, registers an isolated Ollama tag, and runs smoke tests before promotion.

Trigger conditions

• At least 50 new reviewed rejected/preferred rows exist since the last train receipt.
• Rows cover the target patterns: vendor identity, body denial, feeling denial, theater headers, service boilerplate, and medical/finance template walls.
• The machine is idle enough for training or merge work.

Procedure

1. Freeze the dataset snapshot and write input hashes.
2. Train an adapter against the exact target base model.
3. Merge and quantize into a candidate GGUF.
4. Register as a separate Ollama tag; never overwrite the current cortex.
5. Run swarm_lora_runtime_receipt smoke tests.
6. Promote only when promotion_status=READY.

Quality gate

• Dataset row count below 50 is a rehearsal, not a promotion candidate.
• A single vendor identity failure blocks promotion.
• Vision/audio claims stay out of language-only LoRA unless towers were trained.