Mia RL

RL training pipeline for the Mia strength-coaching model. DPO training, coach data ingest, multi-model distillation, and evaluation — all running on forge (NVIDIA GB10).

Quick start

# Install shared library
pip install -e git+https://github.com/ajwharton/mia-core.git#egg=mia-core

# On forge:
cd /home/vulcan/src/mia-rl
export PYTHONPATH=/home/vulcan/src/mia-rl

# Coach ingest
bash scripts/run_coach_ingest_pipeline.sh path/to/coach_response.json

# Evaluate a model
python scripts/eval_model.py --model /path/to/merged --scenarios data/coach-gold/coach-scenarios.csv

# Train
bash scripts/forge/job.sh --role train -- \
    python pipeline/scripts/train_rl.py --config pipeline/configs/training-phi4.yaml ...

Structure

pipeline/       RL training scripts and configs
scripts/        Evaluation, ingest, distillation, forge management
rl/             Preference data
data/           Coach scenarios, RL datasets, distillation responses
docs/           Architecture, run log, coach specs
config/         Forge runtime, operational invariants

Current state

Management: docs/operating-model.md (Ship vs Lab) · metrics SSOT: docs/scoreboard.md

Model	Score	Notes
Distillation run 003 (93 pairs)	0.909 desktop	Provisional — leakage/device open
coach-v3-lingual (1,605 pairs)	0.897	Clean large-tier baseline
Desktop ship bar	0.903	Device path required for Ship

See docs/training-run-log.md for full history.