Upload train_design_sft.py with huggingface_hub

train_design_sft.py

@@ -0,0 +1,98 @@
+# /// script
+# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio", "datasets", "accelerate"]
+# ///
+
+from datasets import load_dataset, concatenate_datasets
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
+import trackio
+
+print("Loading datasets...")
+
+# Load the validated Tailwind dataset (already in messages format)
+tailwind_ds = load_dataset("summykai/tailwind-v4-sft-mix-001", split="train")
+print(f"Tailwind dataset: {len(tailwind_ds)} examples")
+
+# Load the high-quality UI generation dataset
+uigen_ds = load_dataset("smirki/UIGEN-T1.1-TAILWIND", split="train")
+print(f"UIGEN dataset: {len(uigen_ds)} examples")
+
+# Format UIGEN dataset to messages format
+def format_uigen_to_messages(example):
+    # Extract code from markdown code blocks if present
+    answer = example["answer"]
+    if answer.startswith("```html"):
+        answer = answer[7:]  # Remove ```html
+    if answer.endswith("```"):
+        answer = answer[:-3]  # Remove trailing ```
+
+    # Include the reasoning as part of a richer response
+    response = answer.strip()
+
+    return {
+        "messages": [
+            {"role": "user", "content": example["question"]},
+            {"role": "assistant", "content": response}
+        ]
+    }
+
+print("Formatting UIGEN dataset...")
+uigen_formatted = uigen_ds.map(
+    format_uigen_to_messages,
+    remove_columns=uigen_ds.column_names
+)
+
+# Combine datasets
+print("Combining datasets...")
+combined_dataset = concatenate_datasets([tailwind_ds, uigen_formatted])
+combined_dataset = combined_dataset.shuffle(seed=42)
+
+print(f"Total training examples: {len(combined_dataset)}")
+
+# Create train/eval split
+dataset_split = combined_dataset.train_test_split(test_size=0.05, seed=42)
+print(f"Train: {len(dataset_split['train'])}, Eval: {len(dataset_split['test'])}")
+
+print("Initializing trainer...")
+trainer = SFTTrainer(
+    model="chaddy81/qwen3-0.6b-multicode-sft",
+    train_dataset=dataset_split["train"],
+    eval_dataset=dataset_split["test"],
+    peft_config=LoraConfig(
+        r=32,
+        lora_alpha=64,
+        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
+        lora_dropout=0.05,
+        bias="none",
+        task_type="CAUSAL_LM"
+    ),
+    args=SFTConfig(
+        output_dir="qwen3-0.6b-design-sft",
+        push_to_hub=True,
+        hub_model_id="chaddy81/qwen3-0.6b-design-sft",
+        num_train_epochs=3,
+        per_device_train_batch_size=2,
+        gradient_accumulation_steps=8,
+        learning_rate=2e-4,
+        lr_scheduler_type="cosine",
+        warmup_ratio=0.1,
+        max_length=4096,
+        logging_steps=10,
+        eval_strategy="steps",
+        eval_steps=100,
+        save_strategy="steps",
+        save_steps=200,
+        bf16=True,
+        gradient_checkpointing=True,
+        report_to="trackio",
+        run_name="qwen3-design-sft-v1",
+    )
+)
+
+print("Starting training...")
+trainer.train()
+
+print("Pushing to Hub...")
+trainer.push_to_hub()
+
+print("Training complete!")