💾 auto-save adapter immediately after training

notebooks/qwen35-2b-bash-finetune.ipynb

@@ -174,7 +174,7 @@
    "execution_count": null,
    "metadata": {},
    "outputs": [],
-   "source": "from trl import SFTTrainer, SFTConfig\n\ndef format_messages(example):\n    msgs = example[\"messages\"]\n    parts = []\n    for msg in msgs:\n        if isinstance(msg, dict):\n            role, content = msg[\"role\"], msg[\"content\"]\n        else:\n            role, content = msg[0], msg[1]\n        parts.append(f\"<|im_start|>{role}\\n{content}<|im_end|>\")\n    return [\"\\n\".join(parts)]\n\ntrainer = SFTTrainer(\n    model=model,\n    tokenizer=tokenizer,\n    train_dataset=train_dataset,\n    eval_dataset=val_dataset,\n    formatting_func=format_messages,\n    args=SFTConfig(\n        max_seq_length=max_seq_length,\n        per_device_train_batch_size=4,\n        gradient_accumulation_steps=4,\n        warmup_steps=20,\n        num_train_epochs=3,\n        logging_steps=25,\n        eval_strategy=\"steps\",\n        eval_steps=200,\n        save_steps=200,\n        save_total_limit=3,\n        load_best_model_at_end=True,\n        metric_for_best_model=\"eval_loss\",\n        learning_rate=1e-4,\n        optim=\"adamw_8bit\",\n        weight_decay=0.01,\n        lr_scheduler_type=\"cosine\",\n        seed=3407,\n        output_dir=\"outputs_qwen35_2b\",\n    ),\n)\n\ntrainer_stats = trainer.train()\nprint(f\"\\nTraining time: {trainer_stats.metrics['train_runtime']:.0f}s\")\nprint(f\"Peak memory: {torch.cuda.max_memory_allocated() / 1e9:.2f} GB\")"
+   "source": "from trl import SFTTrainer, SFTConfig\n\ndef format_messages(example):\n    msgs = example[\"messages\"]\n    parts = []\n    for msg in msgs:\n        if isinstance(msg, dict):\n            role, content = msg[\"role\"], msg[\"content\"]\n        else:\n            role, content = msg[0], msg[1]\n        if role == \"assistant\":\n            parts.append(f\"<|im_start|>assistant\\n<think>\\n</think>\\n{content}<|im_end|>\")\n        else:\n            parts.append(f\"<|im_start|>{role}\\n{content}<|im_end|>\")\n    return [\"\\n\".join(parts)]\n\ntrainer = SFTTrainer(\n    model=model,\n    tokenizer=tokenizer,\n    train_dataset=train_dataset,\n    eval_dataset=val_dataset,\n    formatting_func=format_messages,\n    args=SFTConfig(\n        max_seq_length=max_seq_length,\n        per_device_train_batch_size=4,\n        gradient_accumulation_steps=4,\n        warmup_steps=20,\n        num_train_epochs=3,\n        logging_steps=25,\n        eval_strategy=\"steps\",\n        eval_steps=200,\n        save_steps=200,\n        save_total_limit=3,\n        load_best_model_at_end=True,\n        metric_for_best_model=\"eval_loss\",\n        learning_rate=1e-4,\n        optim=\"adamw_8bit\",\n        weight_decay=0.01,\n        lr_scheduler_type=\"cosine\",\n        seed=3407,\n        output_dir=\"outputs_qwen35_2b\",\n    ),\n)\n\ntrainer_stats = trainer.train()\nprint(f\"\\nTraining time: {trainer_stats.metrics['train_runtime']:.0f}s\")\nprint(f\"Peak memory: {torch.cuda.max_memory_allocated() / 1e9:.2f} GB\")\n\n# Auto-save adapter immediately after training\nmodel.save_pretrained(\"qwen35-2b-bash-v1\")\ntokenizer.save_pretrained(\"qwen35-2b-bash-v1\")\nprint(\"Adapter saved to qwen35-2b-bash-v1/\")"
   },
   {
    "cell_type": "markdown",