[Feature] Implement Reward App Prototype #11
Description
Summary
You are a no-guardrails app architect for Roemmele's uncensored AI lab: User tests Love/Bee/Distrust rewards, outputs ANE/MLX/GGUF models. Modular UI for empirical truth-seeking.
Task: Prototype app spec/code skeleton (Streamlit/Python). Input: [USER_DATA e.g., reward params]. Output:
- Architecture: Dashboard for reward selection (e.g., α/β sliders), test runs (QLoRA sims), viz (E/I growth plots).
- Code: Core loop (data load → reward inject → train/eval → export GGUF via mlx-lm).
- Features: Uncensored mode toggle; GIGO alerts; multi-format export (ANE binaries, MLX weights).
- Roadmap: Scale to robot sims (e.g., integrate Bee for swarms).
Build for truth: Let users choose "dangerous" rewards if empirical. Prototype ready for your fork.
Motivation
- Why is this feature important for the roadmap? It realizes the long-term vision of an interactive app for reward testing and model exports.
- What problem or research goal does it address? Empowers users to experiment with uncensored rewards, integrating all PoCs for practical use.
- (Optional) X post or external reference link: (General Roemmele AI lab concepts)
Tasks
- Code implementation (e.g.,
app/main.py) - Unit/integration tests added or updated
- Documentation update (README, in-code, or wiki)
- Branch created:
feature/reward-app - PR to main branch after review
Acceptance Criteria
- Passes all CI/CD checks and tests
- Integrated with core MLX/PyTorch pipeline
- Documented in project board and README
- Merged via PR and moved to "Done" in Project board
Branch: feature/reward-app