Upload train_360m_full_sft.py with huggingface_hub

train_360m_full_sft.py

@@ -0,0 +1,93 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#     "trl>=0.12.0",
+#     "peft>=0.7.0",
+#     "transformers>=4.36.0",
+#     "accelerate>=0.24.0",
+#     "trackio",
+#     "datasets",
+# ]
+# ///
+"""
+SFT training for SmolLM2-360M on the FULL 195K recipe dataset.
+Scaling up from 100K to test if more data improves exact match accuracy.
+"""
+
+from datasets import load_dataset
+from peft import LoraConfig
+from trl import SFTTrainer, SFTConfig
+
+# Load the full training set (195K) and validation set (5K)
+train_dataset = load_dataset(
+    "ericlewis/infinite-craft-recipes",
+    data_files="data/train_full.jsonl",
+    split="train",
+)
+eval_dataset = load_dataset(
+    "ericlewis/infinite-craft-recipes",
+    data_files="data/val_5k.jsonl",
+    split="train",
+)
+
+print(f"Train: {len(train_dataset)}, Eval: {len(eval_dataset)}")
+
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=[
+        "q_proj", "k_proj", "v_proj", "o_proj",
+        "gate_proj", "up_proj", "down_proj",
+    ],
+)
+
+config = SFTConfig(
+    output_dir="infinite-craft-smollm2-360m-full",
+    push_to_hub=True,
+    hub_model_id="ericlewis/infinite-craft-smollm2-360m-full",
+    hub_strategy="every_save",
+    hub_private_repo=False,
+
+    # 3 epochs over 195K = ~585K examples
+    num_train_epochs=3,
+    per_device_train_batch_size=32,
+    gradient_accumulation_steps=4,
+    learning_rate=3e-4,
+    max_length=96,
+
+    # Logging & saving
+    logging_steps=50,
+    save_strategy="steps",
+    save_steps=1000,
+    save_total_limit=3,
+
+    # Eval
+    eval_strategy="steps",
+    eval_steps=1000,
+
+    # Schedule
+    warmup_ratio=0.05,
+    lr_scheduler_type="cosine",
+
+    # Precision
+    bf16=True,
+
+    # Tracking
+    report_to="trackio",
+    project="infinite-craft",
+    run_name="smollm2-360m-full-195k-3ep",
+)
+
+trainer = SFTTrainer(
+    model="HuggingFaceTB/SmolLM2-360M-Instruct",
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    peft_config=peft_config,
+    args=config,
+)
+
+trainer.train()
+trainer.push_to_hub()