Upload src/config.yaml with huggingface_hub

src/config.yaml

@@ -0,0 +1,46 @@
+model:
+  name: "answerdotai/ModernBERT-base"
+  loss_function:
+    name: "SentimentWeightedLoss" # Options: "SentimentWeightedLoss", "SentimentFocalLoss"
+    # Parameters for the chosen loss function.
+    # For SentimentFocalLoss, common params are:
+    # gamma_focal: 1.0 # (e.g., 2.0 for standard, -2.0 for reversed, 0 for none)
+    # label_smoothing_epsilon: 0.05 # (e.g., 0.0 to 0.1)
+    # For SentimentWeightedLoss, params is empty:
+    params:
+      gamma_focal: 1.0
+      label_smoothing_epsilon: 0.05
+  output_dir: "checkpoints"
+  max_length: 880 # 256
+  dropout: 0.1
+  # --- Pooling Strategy --- #
+  # Options: "cls", "mean", "cls_mean_concat", "weighted_layer", "cls_weighted_concat"
+  # "cls" uses just the [CLS] token for classification
+  # "mean" uses mean pooling over final hidden states for classification
+  # "cls_mean_concat" uses both [CLS] and mean pooling over final hidden states for classification
+  # "weighted_layer" uses a weighted combination of the final hidden states from the top N layers for classification
+  # "cls_weighted_concat" uses a weighted combination of the final hidden states from the top N layers and the [CLS] token for classification
+  
+  pooling_strategy: "mean" # Current default, change as needed
+
+  num_weighted_layers: 6 # Number of top BERT layers to use for 'weighted_layer' strategies (e.g., 1 to 12 for BERT-base)
+
+data:
+  # No specific data paths needed as we use HF datasets at the moment
+
+training:
+  epochs: 6
+  batch_size: 16
+  lr: 1e-5 # 1e-5 # 2.0e-5
+  weight_decay_rate: 0.02 # 0.01
+  resume_from_checkpoint: "" # "checkpoints/mean_epoch2_0.9361acc_0.9355f1.pt" # Path to checkpoint file, or empty to not resume
+
+inference:
+  # Default path, can be overridden
+  model_path: "checkpoints/mean_epoch5_0.9575acc_0.9575f1.pt" 
+  # Using the same max_length as training for consistency
+  max_length: 880 # 256
+
+
+# "answerdotai/ModernBERT-base"
+# "answerdotai/ModernBERT-large"