# Persona MCQ SDG Context

Use this pack for `sdg/persona_mcq`, which generates persona-grounded
MCQ-shaped **SFT training data**. Use `byob/mcq` instead for a held-out benchmark.

## Contract

- Step: `sdg/persona_mcq`
- Configs: `tiny`, `default`
- Output: `training_jsonl` plus per-teacher blend manifests
- Stages: `personas`, `questions`, `lexical_dedup`, `semantic_dedup`,
  `answer_seed`, `answers`, `build_sft`, `sample`
- Resume key: stable `pipeline.experiment_name` plus unchanged config identity

The pipeline authors questions from configured persona locales, deduplicates,
asks at least three teacher models to answer, votes, applies script/reasoning
quality gates, and exports aligned per-teacher language mixes.

## Routing And Safety

- Start with `tiny` and a fresh experiment name. A reused name with a changed
  config is refused unless overwrite is explicit.
- Teacher agreement is a consistency gate, not factual verification. Inspect
  voting, rejection, and language-purity summaries before training.
- Keep the same teacher list across answer generation, voting, and aligned
  sampling. Distinct aliases should use distinct endpoints in production.
- Language output, persona locale, answer label, script regex, and English
  reasoning contract are independent config choices; do not infer them from a
  language name alone.
- The tiny profile is CPU-portable; production semantic dedup normally uses one
  GPU.

## Credentials

Managed persona assets may require `NGC_API_KEY`. Model endpoints use the
configured base-URL variables and `NVIDIA_API_KEY`; gated Hub models may need
`HF_TOKEN`. Export values in the environment and never write them into YAML,
commands, or committed metadata.

## Downstream Handoff

- AutoModel: feed the selected `training/<teacher>/<view>/train.jsonl` directly
  to `sft/automodel` or `peft/automodel`.
- Megatron-Bridge: feed the emitted blend manifest to
  `data_prep/sft_packing`, then train on the resulting packed Parquet.
- Keep this training corpus separate from BYOB held-out benchmark artifacts.

## Live Sources

- `src/nemotron/steps/sdg/persona_mcq/step.toml`
- `src/nemotron/steps/sdg/persona_mcq/README.md`
- `src/nemotron/steps/sdg/persona_mcq/config/{tiny,default}.yaml`
