Sovereign-Hardened-V1

Multi-benchmark GRPO hardening with 6-signal reward stack and self-healing callback.

Benchmarks Covered

Benchmark	Dataset	Reward Type
GPQA Diamond	Wanfq/gpqa (198 PhD-level MCQs)	Letter-match MCQ
ARC-Challenge	allenai/ai2_arc (1172 MCQs)	Letter-match MCQ
IFEval	google/IFEval (541 IF prompts)	Rule-based constraint verification
PHYBench	Eureka-Lab/PHYBench (500 physics)	Numeric/symbolic match
Adversarial	Generated (injection + IF + tools)	Injection defense + graceful recovery

Self-Healing

Plateau → β×1.5 + LR×0.5 | Collapse → checkpoint + stop

Launch

pip install trl transformers torch datasets accelerate trackio
python benchmark_hardening.py