Upload summaries/evaluation_manifest.json with huggingface_hub
Browse files
summaries/evaluation_manifest.json
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"date": "2026-07-21",
|
| 3 |
+
"data_sha256": "122dfd62dcf8d46ece390ed23335e69cac0d3db8e91479c162d5e4186acdbe0d",
|
| 4 |
+
"split_hash": "2e38193a038079d66cfb5d17dc149af12a9f63a773a2e48ee21d04bb006250fc",
|
| 5 |
+
"split": "direction-invariant polymer SHA1 modulo 2, bucket 0 for evaluation",
|
| 6 |
+
"node_overlap": 0,
|
| 7 |
+
"properties": ["Egc", "Egb"],
|
| 8 |
+
"train_labeled": 290,
|
| 9 |
+
"eval_labeled": 253,
|
| 10 |
+
"train_graph": {"sources": 41, "edges": 271},
|
| 11 |
+
"eval_graph": {"sources": 22, "edges": 48},
|
| 12 |
+
"directions": [[1, 1], [1, -1], [-1, 1], [-1, -1]],
|
| 13 |
+
"preference_weights": 11,
|
| 14 |
+
"candidate_outputs_per_task": 11,
|
| 15 |
+
"edit_budget": 3,
|
| 16 |
+
"eval_tasks": 88,
|
| 17 |
+
"preference_decisions": 968,
|
| 18 |
+
"train_steps": 4321,
|
| 19 |
+
"seeds": [1004, 1005, 1006, 1007, 1008],
|
| 20 |
+
"reward": "R_j=(clip(d_j*(y_j-mu_train,j)/sigma_train,j,-3,3)+3)/6",
|
| 21 |
+
"hypervolume_reference": [0, 0],
|
| 22 |
+
"mip": "mean over source-direction tasks and weights of w^T R",
|
| 23 |
+
"near_optimal_threshold": 0.05,
|
| 24 |
+
"primary_baseline": "greedy_verifier",
|
| 25 |
+
"polybert_revision": "7bf9ed32ac54dea5bc163cf90100728b49341750",
|
| 26 |
+
"eval_graph_clip_rate": 0.0,
|
| 27 |
+
"hardware": "2 local NVIDIA RTX A6000 GPUs; requested B200 endpoints were inaccessible"
|
| 28 |
+
}
|