---
license: apache-2.0
tags:
- grpo
- benchmark-hardening
- adversarial
- gpqa
- ifeval
- physics
base_model:
- Qwen/Qwen2.5-3B-Instruct
datasets:
- Wanfq/gpqa
- allenai/ai2_arc
- google/IFEval
- Eureka-Lab/PHYBench
---

# 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

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