v6-maxlen256: L-2_H-512_A-8 retrained on corrected dataset (label fixes + non-ASCII text cleanups on 3 long articles). Test F1=0.9424, Recall=0.9796, Precision=0.7869, F2=0.9339, threshold=0.16, CPU 30.22ms.

config.json

@@ -3,9 +3,9 @@
     "BertForSequenceClassification"
   ],
   "attention_probs_dropout_prob": 0.1,
-  "classifier_dropout": 0.1893813242387978,
+  "classifier_dropout": 0.10739841773456314,
   "hidden_act": "gelu",
-  "hidden_dropout_prob": 0.1893813242387978,
+  "hidden_dropout_prob": 0.10739841773456314,
   "hidden_size": 512,
   "id2label": {
     "0": "NOT-ACQUISITION",
@@ -24,7 +24,7 @@
   "num_hidden_layers": 2,
   "pad_token_id": 0,
   "position_embedding_type": "absolute",
-  "torch_dtype": "float16",
+  "torch_dtype": "bfloat16",
   "transformers_version": "4.51.1",
   "type_vocab_size": 2,
   "use_cache": true,

experiment_config.json

@@ -1,55 +1,54 @@
 {
   "model": "google/bert_uncased_L-2_H-512_A-8",
-  "best_optuna_trial": 61,
-  "optuna_f1_macro": 0.9488613066118269,
+  "best_optuna_trial": 26,
+  "optuna_f1_macro": 0.93943743058629,
   "hyperparams": {
     "model": "google/bert_uncased_L-2_H-512_A-8",
-    "lr": 8.773852159859744e-05,
+    "lr": 3.646478850567938e-05,
     "batch_size": 8,
-    "scheduler": "cosine_with_min_lr",
-    "min_lr_rate": 0.03975343609380975,
-    "warmup_ratio": 0.17324061161339335,
-    "weight_decay": 0.057521029245153905,
+    "scheduler": "constant_with_warmup",
+    "warmup_ratio": 0.02677317509821073,
+    "weight_decay": 0.3632066548020444,
     "num_epochs": 15,
-    "classifier_dropout": 0.1893813242387978,
-    "label_smoothing": 0.12750721684891234,
+    "classifier_dropout": 0.10739841773456314,
+    "label_smoothing": 0.01230902521971561,
     "use_fgm": true,
-    "fgm_epsilon": 0.32560102802395946
+    "fgm_epsilon": 0.33151445707235533
   },
   "candidates": [
     {
-      "trial": 43,
+      "trial": 29,
       "model": "google/bert_uncased_L-2_H-512_A-8",
-      "optuna_f1": 0.9494,
-      "test_f1_macro": 0.8782
+      "optuna_f1": 0.9403,
+      "test_f1_macro": 0.9415
     },
     {
-      "trial": 11,
+      "trial": 26,
       "model": "google/bert_uncased_L-2_H-512_A-8",
-      "optuna_f1": 0.9493,
-      "test_f1_macro": 0.8694
+      "optuna_f1": 0.9394,
+      "test_f1_macro": 0.9424
     },
     {
-      "trial": 61,
+      "trial": 28,
       "model": "google/bert_uncased_L-2_H-512_A-8",
-      "optuna_f1": 0.9489,
-      "test_f1_macro": 0.8813
+      "optuna_f1": 0.9392,
+      "test_f1_macro": 0.9175
     }
   ],
-  "final_test_f1_macro": 0.8813,
-  "cpu_inference_ms": 24.22,
+  "final_test_f1_macro": 0.9424,
+  "cpu_inference_ms": 30.22,
   "threshold": {
-    "optimal_threshold": 0.26,
-    "f2_score": 0.9057,
-    "precision": 0.6957,
+    "optimal_threshold": 0.16,
+    "f2_score": 0.9339,
+    "precision": 0.7869,
     "recall": 0.9796
   },
   "training_tricks": {
     "loss_type": "ce_weighted",
-    "label_smoothing": 0.12750721684891234,
+    "label_smoothing": 0.01230902521971561,
     "fgm": true,
-    "fgm_epsilon": 0.32560102802395946,
-    "classifier_dropout": 0.1893813242387978,
+    "fgm_epsilon": 0.33151445707235533,
+    "classifier_dropout": 0.10739841773456314,
     "stratified_batch_sampling": true,
     "token_masking": 0.2,
     "token_swapping": 0.15

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bb7b69cb0efb202fa358fda2a02dc761b097f326de23d7fbb35d3062beff1252
-size 44924420
+oid sha256:f0cb3b188f2347f355a6c41c954d97367115675d7101f08ab09c5ae1e2dabea3
+size 44924460

threshold.json

@@ -1,6 +1,6 @@
 {
-  "optimal_threshold": 0.26,
-  "f2_score": 0.9057,
-  "precision": 0.6957,
+  "optimal_threshold": 0.16,
+  "f2_score": 0.9339,
+  "precision": 0.7869,
   "recall": 0.9796
 }

tokenizer.json

@@ -1,7 +1,19 @@
 {
   "version": "1.0",
-  "truncation": null,
-  "padding": null,
+  "truncation": {
+    "direction": "Right",
+    "max_length": 256,
+    "strategy": "LongestFirst",
+    "stride": 0
+  },
+  "padding": {
+    "strategy": "BatchLongest",
+    "direction": "Right",
+    "pad_to_multiple_of": null,
+    "pad_id": 0,
+    "pad_type_id": 0,
+    "pad_token": "[PAD]"
+  },
   "added_tokens": [
     {
       "id": 0,

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4dd73cbca395b028382e67a92b6b9be5de36795224bfbbccc6872b2c8e1b97d2
-size 5368
+oid sha256:59f87e93af7d83772fde74c2c844ce2493e68fcdd33fbbd7ff8c5c92436a7432
+size 5432