Add local training runner and record v1 full training results

Local training infrastructure:
- scripts/run_training.py: CLI training runner with auto-precision
  selection (BF16 → FP32 fallback), preflight checks, structured
  logging, and status.json for programmatic monitoring
- configs/rtx3090.yaml: RTX 3090 config (batch 8 × 4 accum = 32 eff)
- .gitignore: add output/ directory

v1 full training results (A100, BF16, 10 epochs, 135K train examples):
  Overall F1:  0.904 | Precision: 0.907 | Recall: 0.902
  Tier 1 recall: 0.722 (target 0.98) FAIL
  Tier 2 recall: 0.934 (target 0.95) FAIL
  Tier 3 recall: 0.919 (target 0.90) PASS
  Tier 4 recall: 0.866 (target 0.85) PASS

Best checkpoint at epoch 5 (F1=0.903). Epochs 6-7 regressed due to
backbone LR drift (train loss spiked 2.96 → 5.81). load_best_model_at_end
correctly restored epoch 5.

Top types: biometric 0.996, url 0.994, email 0.991, ip_address 0.988
Weak types: pin, passport_number, tax_id, bank_account (rare, <0.70 F1)

WandB: https://wandb.ai/datafog/huggingface/runs/pj8furkq

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: Sid Mohan

.gitignore

@@ -28,6 +28,7 @@ Thumbs.db
 *.onnx
 *.safetensors
 wandb/
+output/
 outputs/
 checkpoints/
 runs/

pii-ner-v1/configs/rtx3090.yaml

@@ -0,0 +1,35 @@
+model:
+  backbone: microsoft/deberta-v3-xsmall
+  char_embed_dim: 50
+  char_vocab_size: 256
+  char_cnn_filters: [50, 50, 50]
+  char_cnn_widths: [3, 4, 5]
+  max_char_len: 20
+  dropout: 0.1
+
+data:
+  max_seq_len: 256
+  val_ratio: 0.1
+  test_ratio: 0.1
+  seed: 42
+
+training:
+  epochs: 10
+  batch_size: 8
+  gradient_accumulation_steps: 4  # effective batch = 32
+  lr_backbone: 2.0e-5
+  lr_head: 1.0e-3
+  warmup_ratio: 0.1
+  weight_decay: 0.01
+  eval_strategy: epoch
+  save_strategy: epoch
+  metric_for_best_model: overall_f1
+  logging_steps: 50
+  dataloader_num_workers: 4
+  save_total_limit: 3
+  output_dir: output
+  run_name: pii-ner-v1-rtx3090
+
+wandb:
+  enabled: true
+  project: datafog-pii-ner