ArsSocratica/egora-benchmarks

---
license: agpl-3.0
task_categories:
- text-generation
tags:
- egora
- lora
- fine-tuning
- rotation-retention-law
- knowledge-retention
- benchmark
pretty_name: EgoRA Benchmark Results
size_categories:
- 1K<n<10K
---

# EgoRA Benchmark Results

Comprehensive benchmark results for **EgoRA** (Entropy-Governed Orthogonality Regularization for Adaptation) across multiple model scales, domains, and architectures.

📦 **Package:** [egora on PyPI](https://pypi.org/project/egora/)
💻 **Code:** [ArsSocratica/EgoRA on GitHub](https://github.com/ArsSocratica/EgoRA)
📄 **Paper:** [arXiv:2602.05192](https://arxiv.org/abs/2602.05192)
🔖 **DOI:** [10.5281/zenodo.19398709](https://doi.org/10.5281/zenodo.19398709)

## Dataset Structure

### `llama-3.2-1b/`, `llama-3.2-3b/`, `llama-3.1-8b/`
Fine-tuning results across 3 model scales, 2 domains (Alpaca general, Medical), 4 methods:
- **Baseline LoRA** — standard low-rank adaptation
- **DoRA** — weight-decomposed low-rank adaptation
- **EgoRA e²** — with entropy-squared governor
- **EgoRA adaptive v2** — with adaptive entropy governor

Each config contains:
- `*_results.json` — benchmark scores (MMLU, TruthfulQA, HellaSwag, Winogrande, MedQA, MedMCQA)
- `history.json` — per-step training curves (loss, penalty, λ, entropy)
- `summary.json` — final metrics summary
- `rotation_analysis_*.json` — per-head rotation geometry

### `threshold-analysis/`
Rotation-Retention Law validation:
- `golden_ratio_k.json` — k-value (ΔM/θ̄) per condition
- `dimensionality_threshold.json` — θ_crit = arcsin(1/√d_head)
- `cross_architecture_threshold.json` — cross-architecture analysis
- `phase_transition.json` — phase transition at θ̄ ≈ 5°

### `cross-modal/`
Cross-modal experiments (Mistral-7B, Phi-3 Mini):
rotation geometry, benchmarks, knowledge maps.

### `figures/`
Publication-ready figures.

## Key Results

| Scale | Method | θ̄ (°) | MMLU Δ (pp) | Damaged Heads |
|-------|--------|--------|-------------|---------------|
| 1B | Baseline LoRA | 6.89 | −3.11 | 51.2% |
| 1B | EgoRA | 2.33 | −1.04 | 2.4% |
| 3B | Baseline LoRA | 6.09 | −4.61 | 49.4% |
| 3B | EgoRA | 2.07 | −1.89 | 1.3% |
| 8B | Baseline LoRA | 7.00 | −3.56 | 56.1% |
| 8B | EgoRA | 2.57 | −0.56 | 3.8% |

**Rotation-Retention Law:** θ̄ > 5° → MMLU loss > 3pp; θ̄ < 3° → MMLU loss < 2pp.

## Citation

```bibtex
@article{dillerop2026egora,
title={EgoRA: Entropy-Governed Orthogonality Regularization for Adaptation},
author={Dillerop, Mark},
journal={arXiv preprint arXiv:2602.05192},
year={2026}
}
```

## License

AGPL-3.0 with Academic Additional Permission (Section 7).
Academic use is free with citation. Commercial use requires a separate license.
Contact: [e-mail]