Swastikr/polyglot-optima-openenv
Polyglot-Optima
Polyglot-Optima is an OpenEnv environment for training an LLM to translate Python functions into hardware-aware C++ that is both fast and correct.
Problem
LLMs can generate optimized code, but often fail on edge-case correctness, portability, and anti-gaming behavior (fast but wrong outputs). This environment targets that gap with closed-loop tool use and verifiable rewards.
Environment Design
- API shape: Gym-style
reset,step,state. - 3-round episodes: iterative refinement, final submission at round 3.
- 9 tools: profiling, complexity analysis, memory analysis, compile+benchmark, equivalence verifier, portability checker, and final submit.
- Reward DAG: composable rubrics for speedup, correctness, diagnosis quality, portability, and self-correction.
- Continuous rewards: no hard 0/1 optimization cliff in the main learning path.
Innovation Highlights
- Adaptive 4-axis curriculum updates global difficulty over batches.
- Adversarial trap library with category-focused adaptive resampling from recent failures.
- Semantic trap variation (AST-level no-op rewrites) to reduce memorization.
- Roofline-aware speedup scoring for hardware-grounded performance reward.
- Anti-gaming verification through fuzzing + adversarial pass checks.
Why This Matters
The target behavior is not just "compile and run", but robust optimization under realistic constraints: correctness under adversarial inputs, reasoning about bottlenecks, and hardware-aware strategy selection.
Local Usage
python -m pytest -q
python -m ruff check .Run smoke LLM integration:
python tests/smoke_llm_hf.pyCursor/OpenAI-compatible provider mode:
export LLM_PROVIDER=cursor
export CURSOR_API_KEY=...
export CURSOR_MODEL=gpt-4.1-nano
python tests/smoke_llm_hf.pyNotebook Usage and HF Spaces
You can use this environment directly in a local notebook without deploying to HF Spaces.
- For development/training: local usage is enough.
- For hackathon submission: deploy to HF Spaces and link it in README per requirements.
Current Validation Snapshot
- Unit/integration tests passing.
- Smoke integration path validates parseability/tool-loop behavior.
- Reward and gate tests verify coherent scoring behavior.
Results (Judge-facing)
After running training/openenv_hackathon_training.ipynb, add:
- Reward distribution plot:
docs/plots/reward_distribution_baseline_vs_trained.png - Correctness curve plot:
docs/plots/correctness_baseline_vs_trained.png - Baseline vs trained metrics table (reward mean, correctness, compile rate, portability).
Required Submission Links
Add these links before final submission:
- HF Space (environment URL judges will pull):
TODO_ADD_HF_SPACE_URL - Training notebook/script:
training/openenv_hackathon_training.ipynb - W&B run (or equivalent training evidence):
TODO_ADD_WANDB_RUN_URL - Short writeup/video/slides (<2 min video or mini blog):
TODO_ADD_STORY_URL
Submission Checklist (from hackathon PDF)
- [ ] Environment deployed to HF Space and URL added above
- [x] Valid OpenEnv manifest (
openenv.yaml) present - [x] Training notebook/script using TRL/Unsloth path present
- [ ] Real training evidence linked (loss/reward curves from an actual run)
- [ ] README includes all judge-facing links (Space + writeup/video/slides + run logs)
- [ ] Key plots embedded and committed in repo (
docs/plots/*.png)
