Upload train.py with huggingface_hub

train.py

@@ -0,0 +1,62 @@
+# /// script
+# dependencies = [
+#   "trl>=0.12.0",
+#   "peft>=0.13.0",
+#   "transformers>=4.45.0",
+#   "datasets>=3.0.0",
+#   "accelerate>=1.0.0",
+#   "trackio",
+# ]
+# ///
+"""Lexwell SFT — Qwen2.5-3B-Instruct fine-tune on IRAC contract-review corpus.
+
+Runs on HF Jobs. Reads 80 hand-curated SFT rows from Curious-PM/lexwell-contract-irac
+and pushes the LoRA adapter to Curious-PM/lexwell-contract-irac-qwen2.5-3b-lora.
+"""
+from datasets import load_dataset
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
+
+BASE  = "Qwen/Qwen2.5-3B-Instruct"
+DS    = "Curious-PM/lexwell-contract-irac"
+OUT   = "Curious-PM/lexwell-contract-irac-qwen2.5-3b-lora"
+
+print(f"Loading dataset {DS}...")
+ds = load_dataset(DS, data_files="lexwell_v2.jsonl", split="train")
+print(f"Loaded {len(ds)} rows")
+print(f"First row keys: {list(ds[0].keys())}")
+
+trainer = SFTTrainer(
+    model=BASE,
+    train_dataset=ds,
+    peft_config=LoraConfig(
+        r=16,
+        lora_alpha=32,
+        task_type="CAUSAL_LM",
+        target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
+    ),
+    args=SFTConfig(
+        output_dir="lexwell-lora",
+        num_train_epochs=10,
+        per_device_train_batch_size=4,
+        gradient_accumulation_steps=1,
+        learning_rate=2e-4,
+        max_length=2048,
+        logging_steps=10,
+        eval_strategy="no",
+        save_strategy="no",
+        bf16=True,
+        push_to_hub=True,
+        hub_model_id=OUT,
+        hub_strategy="end",
+        report_to="trackio",
+        project="lexwell-irac",
+        run_name="qwen2.5-3b-sft-r16-e10",
+    ),
+)
+
+print("Starting training...")
+trainer.train()
+print("Training complete. Pushing to Hub...")
+trainer.push_to_hub()
+print(f"\nAdapter pushed: https://huggingface.co/{OUT}")