add20k retrain (seed 44): 20k corpus negatives, 24.0% train prevalence; PubMed fire rate 19.46% -> 0.51%; test F1 0.914

config.json

@@ -20,9 +20,17 @@
   "gradient_checkpointing": false,
   "hidden_activation": "gelu",
   "hidden_size": 1024,
+  "id2label": {
+    "0": "NOT_RELEVANT",
+    "1": "RELEVANT"
+  },
   "initializer_cutoff_factor": 2.0,
   "initializer_range": 0.02,
   "intermediate_size": 2624,
+  "label2id": {
+    "NOT_RELEVANT": 0,
+    "RELEVANT": 1
+  },
   "layer_norm_eps": 1e-05,
   "local_attention": 128,
   "local_rope_theta": 10000.0,
@@ -44,13 +52,5 @@
   "sparse_prediction": false,
   "torch_dtype": "float32",
   "transformers_version": "4.48.3",
-  "vocab_size": 50368,
-  "id2label": {
-    "0": "NOT_RELEVANT",
-    "1": "RELEVANT"
-  },
-  "label2id": {
-    "NOT_RELEVANT": 0,
-    "RELEVANT": 1
-  }
-}
+  "vocab_size": 50368
+}

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c55cf6de446d0080f2447a638e01cb7b8be1dda78e9ec7c9dc3671a32f2963be
+oid sha256:6b868b0bbd123983036685ffb595d00a3220f1b096bd918e69884f2c7a4a9c5f
 size 1583351632