code/self_healing_finetuning/prompt_packs/model_selection.md

# Model Selection Prompt Pack

Apply SHFT model-profile policy. Do not hardcode a single base model into launchers, providers, trainers, proof jobs, or generated documentation.

Default profile:

- `fingpt`
- candidate: `linvest21/linvest21_fingpt_v1_000`
- base model: `meta-llama/Meta-Llama-3-8B`
- start behavior: approved Linvest21 FinGPT adapter bootstrap
- source FinGPT adapter lineage: `FinGPT/fingpt-mt_llama3-8b_lora`
- license posture: Meta Llama 3 community license plus FinGPT adapter terms; commercial review required

Open commercial foundation profile:

- `qwen3_32b`
- candidate/base model: `Qwen/Qwen3-32B`
- start behavior: fresh QLoRA from the base model, no baseline adapter
- proof baseline: raw `Qwen/Qwen3-32B`
- license posture: Apache-2.0; commercial use allowed

Operator controls:

```bat
python -m n21.cli select-model --task finance_qa --env dev --model-profile fingpt
python -m n21.cli select-model --task finance_qa --env dev --model-profile qwen3_32b
set SHFT_MODEL_PROFILE=qwen3_32b
```

Implementation rules:

- Read profile definitions from `configs/model_profiles.json`.
- Resolve profiles through `model_policy/profiles.py`.
- Preserve `SHFT_MODEL_CANDIDATE` and `SHFT_BASE_MODEL_ID` only as explicit overrides.
- Keep paired proof baseline-aware: adapter baseline for `fingpt`, raw base-model baseline for `qwen3_32b`.
- Adding a future model should mean adding a profile entry, not editing every script.