egora-benchmarks / threshold-analysis /cross_architecture_threshold.json
ArsSocratica
Upload EgoRA benchmark results: Llama 1B/3B/8B (Alpaca+Medical), cross-modal (Mistral-7B, Phi-3 Mini, LLaVA-7B, Qwen2-VL-7B), rotation geometry, threshold analysis
bf34f0d
{
"n_models": 5,
"models": [
"1B_medical",
"3B_medical",
"1B_alpaca",
"3B_alpaca",
"8B_alpaca"
],
"predictions_deg": [
7.1808,
5.0709,
7.1808,
5.0709,
5.0709
],
"empiricals_deg": [
7.0871,
6.0861,
6.7668,
6.046,
7.3874
],
"correlation": 0.3854,
"mae_deg": 0.9629,
"formula_confirmed": false
}