Upload training_config.yaml with huggingface_hub

training_config.yaml

@@ -0,0 +1,30 @@
+model:
+  base_model: BAAI/bge-reranker-v2-m3
+  freeze_encoder: true
+  max_length: 8192
+  max_sentences: 500  # NEW: Prevent OOM/hang on very long docs
+  pruning_head:
+    hidden_dim: 512
+    dropout: 0.2
+training:
+  labels_file: /workspace/data/labels/examples.jsonl
+  train_split: 0.9
+  epochs: 10
+  batch_size: 8  # Reduced from 16 for long docs
+  gradient_accumulation: 8  # Maintain effective batch 64
+  learning_rate: 0.0001  # Slightly lower
+  weight_decay: 0.02
+  warmup_steps: 200  # More warmup
+  max_grad_norm: 1.0
+loss:
+  loss_type: bce
+  pos_weight: 70.0  # FIXED: Was 20, now matches 94:1 imbalance
+  label_smoothing: 0.05
+hardware:
+  mixed_precision: bf16
+wandb:
+  enabled: false
+global:
+  data_dir: /workspace/data
+  checkpoint_dir: /workspace/checkpoints/pruner_optimized
+  log_file: /workspace/logs/pruner_sentence_v2.log