feat: add CLI runner + training recipe docs

ai-ml/hf-finetuning/run_finetune.py

@@ -0,0 +1,67 @@
+# =============================================================================
+# SFT Fine-Tuning — CLI Entry Point (LoRA Without Regret config)
+# =============================================================================
+# Usage:
+#   # Default: tulu-3-sft + Llama-3.1-8B
+#   python run_finetune.py
+#
+#   # OpenThoughts reasoning dataset
+#   python run_finetune.py --dataset_key openthoughts-114k
+#
+#   # Ultrachat fallback
+#   python run_finetune.py --dataset_key ultrachat-200k
+#
+#   # Custom hub model ID
+#   python run_finetune.py --hub_model_id my-org/my-model-v2
+# =============================================================================
+
+import argparse
+import sys
+from finetune import FinetuneConfig, finetune, DATASET_REGISTRY
+
+
+def main():
+    parser = argparse.ArgumentParser(description="SFT Fine-Tuning (LoRA Without Regret)")
+    parser.add_argument("--dataset_key", default="tulu-3-sft",
+                        choices=list(DATASET_REGISTRY.keys()),
+                        help="Dataset to train on")
+    parser.add_argument("--hub_model_id", default=None,
+                        help="HuggingFace Hub model ID for push")
+    parser.add_argument("--num_train_epochs", type=int, default=None)
+    parser.add_argument("--learning_rate", type=float, default=None)
+    parser.add_argument("--lora_r", type=int, default=None)
+    parser.add_argument("--per_device_train_batch_size", type=int, default=None)
+    parser.add_argument("--max_seq_length", type=int, default=None)
+
+    args = parser.parse_args()
+
+    config = FinetuneConfig()
+    if args.dataset_key:
+        config.dataset_key = args.dataset_key
+    if args.hub_model_id:
+        config.hub_model_id = args.hub_model_id
+    if args.num_train_epochs:
+        config.num_train_epochs = args.num_train_epochs
+    if args.learning_rate:
+        config.learning_rate = args.learning_rate
+    if args.lora_r:
+        config.lora_r = args.lora_r
+    if args.per_device_train_batch_size:
+        config.per_device_train_batch_size = args.per_device_train_batch_size
+    if args.max_seq_length:
+        config.max_seq_length = args.max_seq_length
+
+    print(f"Config: model={config.model_name}")
+    print(f"  dataset={config.dataset_key}")
+    print(f"  lora_r={config.lora_r}, lora_alpha={config.lora_alpha}")
+    print(f"  target_modules={config.target_modules}")
+    print(f"  lr={config.learning_rate}, epochs={config.num_train_epochs}")
+    print(f"  effective_batch={config.per_device_train_batch_size * config.gradient_accumulation_steps}")
+    print(f"  packing={config.packing}, strategy={config.packing_strategy}")
+    print(f"  assistant_only_loss={config.assistant_only_loss}")
+
+    finetune(config)
+
+
+if __name__ == "__main__":
+    main()