# TenaOS LoRA Training Code

This directory contains the training, merge, conversion, and evaluation helper
scripts used for the TenaOS task-tagged LoRA release.

The scripts are provided for reproducibility and auditability. They are not
required to run the TenaOS clinical demo.

## Main Scripts

| Script | Purpose |
| --- | --- |
| `prepare_lora_corpus.py` | Normalize and prepare task-tagged SFT records |
| `train_unsloth_lora.py` | Train the LoRA adapter with Unsloth/TRL |
| `merge_lora_checkpoint.py` | Merge a LoRA checkpoint into BF16 base weights |
| `convert_merged_to_gguf.py` | Convert merged Hugging Face weights to GGUF |
| `evaluate_lora_ab.py` | Sidecar base-vs-LoRA smoke/eval harness |
| `evaluate_lora_ab_v2.py` | Tool-call focused held-out eval harness |
| `check_chat_template_parity.py` | Check tokenizer/chat-template consistency |
| `reconstruct_kb_hits.py` | Reconstruct KB-hit metadata used by traces |