eval artifact rerank_eval.json

eval/rerank_eval.json

@@ -0,0 +1,40 @@
+{
+  "backbone": "meta-llama/Llama-3.2-3B",
+  "layer": 20,
+  "av_ckpt": "artifacts/nla/llama32_3B/sft/best_av.pt",
+  "M": 200,
+  "K": 32,
+  "T": 64,
+  "temperature": 1.0,
+  "pool_size": 2000,
+  "anisotropy_mu_norm": 7.211436748504639,
+  "greedy_cen_mean": 0.619103729724884,
+  "greedy_raw_mean": 0.6656057238578796,
+  "sampled_cen_mean": 0.6248369812965393,
+  "sampled_raw_mean": 0.6734464764595032,
+  "kcurve_oracle_top1_cen": {
+    "1": 0.6356024742126465,
+    "2": 0.6775426268577576,
+    "4": 0.716196596622467,
+    "8": 0.7475207448005676,
+    "16": 0.7802464365959167,
+    "32": 0.8092434406280518
+  },
+  "kcurve_oracle_top1_raw": {
+    "1": 0.6818430423736572,
+    "2": 0.7163781523704529,
+    "4": 0.7506033182144165,
+    "8": 0.7772679328918457,
+    "16": 0.8067382574081421,
+    "32": 0.8324524164199829
+  },
+  "logp_rerank_cen_mean": 0.6240078210830688,
+  "logp_rerank_raw_mean": 0.6711649298667908,
+  "logp_rerank_vs_greedy": 0.0049040913581848145,
+  "nn_anchor_rerank_cen_mean": 0.783254861831665,
+  "nn_anchor_baseline_cen_mean": 0.8357419371604919,
+  "spearman_logp_vs_oracle_cen_mean": 0.05462426692247391,
+  "spearman_logp_vs_oracle_cen_p50": 0.059017594903707504,
+  "spearman_logp_vs_oracle_cen_p05": -0.4031524956226349,
+  "spearman_logp_vs_oracle_cen_p95": 0.5307551622390747
+}