    generative_ai/sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "af2ee7fb-e888-4e38-a349-c7c40dfd2963",
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "source": [
    "# Fine-tune LLaMA 3.2 models on SageMaker JumpStart"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0679163a-387a-4ba6-8ce0-d5571614c0dc",
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "source": [
    "---\n",
    "\n",
    "This notebook's CI test result for us-west-2 is as follows. CI test results in other regions can be found at the end of the notebook.\n",
    "\n",
    "![This us-west-2 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/us-west-2/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "251624f9-1eb6-4051-a774-0a4ba83cabf5",
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "source": [
    "---\n",
    "In this demo notebook, we demonstrate how to use the SageMaker Python SDK to deploy pre-trained Llama 3.2 model as well as fine-tune it for your dataset in domain adaptation or instruction tuning format.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e3d9b99d-639b-40f3-91e3-1fe00ee032a4",
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "source": [
    "### Model License information\n",
    "---\n",
    "To perform inference on these models, you need to pass custom_attributes='accept_eula=true' as part of header. This means you have read and accept the end-user-license-agreement (EULA) of the model. EULA can be found in model card description or from https://ai.meta.com/resources/models-and-libraries/llama-downloads/. By default, this notebook sets custom_attributes='accept_eula=false', so all inference requests will fail until you explicitly change this custom attribute.\n",
    "\n",
    "Note: Custom_attributes used to pass EULA are key/value pairs. The key and value are separated by '=' and pairs are separated by ';'. If the user passes the same key more than once, the last value is kept and passed to the script handler (i.e., in this case, used for conditional logic). For example, if 'accept_eula=false; accept_eula=true' is passed to the server, then 'accept_eula=true' is kept and passed to the script handler.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "019c4fcd-d6c5-4381-8425-1d224c0ac197",
   "metadata": {
    "collapsed": true,
    "jupyter": {
     "outputs_hidden": true
    },
    "tags": []
   },
   "source": [
    "### Set up\n",
    "\n",
    "---\n",
    "We begin by installing and upgrading necessary packages. Restart the kernel after executing the cell below for the first time.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "85addd9d-ec89-44a7-9fb5-9bc24fe9993b",
   "metadata": {
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "!pip install --upgrade sagemaker datasets"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "13274b9b-87bd-4090-a6aa-294570c31e0e",
   "metadata": {},
   "source": [
    "## Deploy Pre-trained Model\n",
    "\n",
    "---\n",
    "\n",
    "First we will deploy the Llama 3.2 model as a SageMaker endpoint. To train/deploy 1B and 2B models, please change model_id to \"meta-textgeneration-llama-3-2-1b\" and \"meta-textgeneration-llama-3-2-3b\" respectively.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c7e01401-82db-4d49-9457-f930f4138618",
   "metadata": {
    "jumpStartAlterations": [
     "modelIdVersion"
    ],
    "tags": []
   },
   "outputs": [],
   "source": [
    "model_id, model_version = \"meta-textgeneration-llama-3-2-3b\", \"1.*\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1722b230-b7bc-487f-b4ee-98ca42848423",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "from sagemaker.jumpstart.model import JumpStartModel\n",
    "\n",
    "pretrained_model = JumpStartModel(model_id=model_id, model_version=model_version)\n",
    "# Please change the following line to have accept_eula = True\n",
    "pretrained_predictor = pretrained_model.deploy(accept_eula=False)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8017c4ef-eb89-4da6-8e28-c800adbfc4b8",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Invoke the endpoint\n",
    "\n",
    "---\n",
    "Next, we invoke the endpoint with some sample queries. Later, in this notebook, we will fine-tune this model with a custom dataset and carry out inference using the fine-tuned model. We will also show comparison between results obtained via the pre-trained and the fine-tuned models.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b795a085-048f-42b2-945f-0cd339c1cf91",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "def print_response(payload, response):\n",
    "    print(payload[\"inputs\"])\n",
    "    print(f\"> {response.get('generated_text')}\")\n",
    "    print(\"\\n==================================\\n\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5dd833f8-1ddc-4805-80b2-19e7db629880",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "payload = {\n",
    "    \"inputs\": \"I believe the meaning of life is\",\n",
    "    \"parameters\": {\n",
    "        \"max_new_tokens\": 64,\n",
    "        \"top_p\": 0.9,\n",
    "        \"temperature\": 0.6,\n",
    "        \"return_full_text\": False,\n",
    "    },\n",
    "}\n",
    "try:\n",
    "    response = pretrained_predictor.predict(\n",
    "        payload, custom_attributes=\"accept_eula=false\"  # Please change this to \"accept_eula=true\"\n",
    "    )\n",
    "    print_response(payload, response)\n",
    "except Exception as e:\n",
    "    print(e)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5e19e16f-d459-40c6-9d6b-0272938b3878",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Dataset preparation for fine-tuning\n",
    "\n",
    "---\n",
    "\n",
    "You can fine-tune on the dataset with domain adaptation format or instruction tuning format. Please find more details in the section [Dataset instruction](#Dataset-instruction). In this demo, we will use a subset of [Dolly dataset](https://huggingface.co/datasets/databricks/databricks-dolly-15k) in an instruction tuning format. Dolly dataset contains roughly 15,000 instruction following records for various categories such as question answering, summarization, information extraction etc. It is available under Apache 2.0 license. We will select the summarization examples for fine-tuning.\n",
    "\n",
    "\n",
    "Training data is formatted in JSON lines (.jsonl) format, where each line is a dictionary representing a single data sample. All training data must be in a single folder, however it can be saved in multiple jsonl files. The training folder can also contain a template.json file describing the input and output formats.\n",
    "\n",
    "To train your model on a collection of unstructured dataset (text files), please see the section [Example fine-tuning with Domain-Adaptation dataset format](#Example-fine-tuning-with-Domain-Adaptation-dataset-format) in the Appendix.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6dd20a0d-15a5-49b0-a330-a75755d046ed",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "from datasets import load_dataset\n",
    "\n",
    "dolly_dataset = load_dataset(\"databricks/databricks-dolly-15k\", split=\"train\")\n",
    "\n",
    "# To train for question answering/information extraction, you can replace the assertion in next line to example[\"category\"] == \"closed_qa\"/\"information_extraction\".\n",
    "summarization_dataset = dolly_dataset.filter(lambda example: example[\"category\"] == \"summarization\")\n",
    "summarization_dataset = summarization_dataset.remove_columns(\"category\")\n",
    "\n",
    "# We split the dataset into two where test data is used to evaluate at the end.\n",
    "train_and_test_dataset = summarization_dataset.train_test_split(test_size=0.1)\n",
    "\n",
    "# Dumping the training data to a local file to be used for training.\n",
    "train_and_test_dataset[\"train\"].to_json(\"train.jsonl\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e9fbf002-3ee3-4cc8-8fce-871939f1bd19",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "train_and_test_dataset[\"train\"][0]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9b2e5489-33dc-4623-92da-f6fc97bd25ab",
   "metadata": {},
   "source": [
    "---\n",
    "Next, we create a prompt template for using the data in an instruction / input format for the training job (since we are instruction fine-tuning the model in this example), and also for inferencing the deployed endpoint.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "90451114-7cf5-445c-88e3-02ccaa5d3a4b",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "import json\n",
    "\n",
    "template = {\n",
    "    \"prompt\": \"Below is an instruction that describes a task, paired with an input that provides further context. \"\n",
    "    \"Write a response that appropriately completes the request.\\n\\n\"\n",
    "    \"### Instruction:\\n{instruction}\\n\\n### Input:\\n{context}\\n\\n\",\n",
    "    \"completion\": \" {response}\",\n",
    "}\n",
    "with open(\"template.json\", \"w\") as f:\n",
    "    json.dump(template, f)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a22171b1-1cec-4cec-9ce4-db62761633d9",
   "metadata": {
    "tags": []
   },
   "source": [
    "### Upload dataset to S3\n",
    "---\n",
    "\n",
    "We will upload the prepared dataset to S3 which will be used for fine-tuning.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5e1ee29a-8439-4788-8088-35a433fe2110",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "from sagemaker.s3 import S3Uploader\n",
    "import sagemaker\n",
    "import random\n",
    "\n",
    "output_bucket = sagemaker.Session().default_bucket()\n",
    "default_bucket_prefix = sagemaker.Session().default_bucket_prefix\n",
    "default_bucket_prefix_path = \"\"\n",
    "\n",
    "# If a default bucket prefix is specified, append it to the s3 path\n",
    "if default_bucket_prefix:\n",
    "    default_bucket_prefix_path = f\"/{default_bucket_prefix}\"\n",
    "\n",
    "local_data_file = \"train.jsonl\"\n",
    "train_data_location = f\"s3://{output_bucket}{default_bucket_prefix_path}/dolly_dataset\"\n",
    "S3Uploader.upload(local_data_file, train_data_location)\n",
    "S3Uploader.upload(\"template.json\", train_data_location)\n",
    "print(f\"Training data: {train_data_location}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "86e61340-bc81-477d-aaf1-f37e8c554863",
   "metadata": {
    "tags": []
   },
   "source": [
    "## Train the model\n",
    "---\n",
    "Next, we fine-tune the LLaMA 3.2 3B model on the summarization dataset from Dolly. Finetuning scripts are based on scripts provided by [this repo](https://github.com/meta-llama/llama-recipes/tree/main). To learn more about the fine-tuning scripts, please checkout section [5. Few notes about the fine-tuning method](#5.-Few-notes-about-the-fine-tuning-method). For a list of supported hyper-parameters and their default values, please see section [3. Supported Hyper-parameters for fine-tuning](#3.-Supported-Hyper-parameters-for-fine-tuning).\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9a71087e-9c9e-42d7-999e-5f3fac07bc4a",
   "metadata": {
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "from sagemaker.jumpstart.estimator import JumpStartEstimator\n",
    "\n",
    "\n",
    "estimator = JumpStartEstimator(\n",
    "    model_id=model_id,\n",
    "    model_version=model_version,\n",
    "    environment={\"accept_eula\": \"false\"},  # Please change this to {\"accept_eula\": \"true\"}\n",
    "    disable_output_compression=True,\n",
    ")\n",
    "# By default, instruction tuning is set to false. Thus, to use instruction tuning dataset you use\n",
    "estimator.set_hyperparameters(instruction_tuned=\"True\", epoch=\"5\", max_input_length=\"1024\")\n",
    "estimator.fit({\"training\": train_data_location})"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3e3889d9-1567-41ad-9375-fb738db629fa",
   "metadata": {
    "tags": []
   },
   "source": [
    "Studio Kernel Dying issue:  If your studio kernel dies and you lose reference to the estimator object, please see section [6. Studio Kernel Dead/Creating JumpStart Model from the training Job](#6.-Studio-Kernel-Dead/Creating-JumpStart-Model-from-the-training-Job) on how to deploy endpoint using the training job name and the model id. \n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5e9decbf-08c6-4cb4-8644-4a96afb5bebf",
   "metadata": {
    "tags": []
   },
   "source": [
    "### Deploy the fine-tuned model\n",
    "---\n",
    "Next, we deploy fine-tuned model. We will compare the performance of fine-tuned and pre-trained model.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "016e591b-63f8-4e0f-941c-4b4e0b9dc6fc",
   "metadata": {},
   "outputs": [],
   "source": [
    "finetuned_predictor = estimator.deploy()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "cb57904a-9631-45fe-bc3f-ae2fbb992960",
   "metadata": {
    "tags": []
   },
   "source": [
    "### Evaluate the pre-trained and fine-tuned model\n",
    "---\n",
    "Next, we use the test data to evaluate the performance of the fine-tuned model and compare it with the pre-trained model. \n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "87085bf6-dc7e-46f3-8563-d2e4aafd0820",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "from IPython.display import display, HTML\n",
    "\n",
    "test_dataset = train_and_test_dataset[\"test\"]\n",
    "\n",
    "(\n",
    "    inputs,\n",
    "    ground_truth_responses,\n",
    "    responses_before_finetuning,\n",
    "    responses_after_finetuning,\n",
    ") = (\n",
    "    [],\n",
    "    [],\n",
    "    [],\n",
    "    [],\n",
    ")\n",
    "\n",
    "\n",
    "def predict_and_print(datapoint):\n",
    "    # For instruction fine-tuning, we insert a special key between input and output\n",
    "    input_output_demarkation_key = \"\\n\\n### Response:\\n\"\n",
    "\n",
    "    payload = {\n",
    "        \"inputs\": template[\"prompt\"].format(\n",
    "            instruction=datapoint[\"instruction\"], context=datapoint[\"context\"]\n",
    "        )\n",
    "        + input_output_demarkation_key,\n",
    "        \"parameters\": {\"max_new_tokens\": 100},\n",
    "    }\n",
    "    inputs.append(payload[\"inputs\"])\n",
    "    ground_truth_responses.append(datapoint[\"response\"])\n",
    "    # Please change the following line to \"accept_eula=true\"\n",
    "    pretrained_response = pretrained_predictor.predict(\n",
    "        payload, custom_attributes=\"accept_eula=false\"\n",
    "    )\n",
    "    responses_before_finetuning.append(pretrained_response.get(\"generated_text\"))\n",
    "    # Fine Tuned Llama 3.2models doesn't required to set \"accept_eula=true\"\n",
    "    finetuned_response = finetuned_predictor.predict(payload)\n",
    "    responses_after_finetuning.append(finetuned_response.get(\"generated_text\"))\n",
    "\n",
    "\n",
    "try:\n",
    "    for i, datapoint in enumerate(test_dataset.select(range(5))):\n",
    "        predict_and_print(datapoint)\n",
    "\n",
    "    df = pd.DataFrame(\n",
    "        {\n",
    "            \"Inputs\": inputs,\n",
    "            \"Ground Truth\": ground_truth_responses,\n",
    "            \"Response from non-finetuned model\": responses_before_finetuning,\n",
    "            \"Response from fine-tuned model\": responses_after_finetuning,\n",
    "        }\n",
    "    )\n",
    "    display(HTML(df.to_html()))\n",
    "except Exception as e:\n",
    "    print(e)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f6b0a0f5-ef34-40db-8ab7-c24a5d14b525",
   "metadata": {},
   "source": [
    "### Clean up resources"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d73ab2da-d00f-46db-90eb-81812898653b",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Delete resources\n",
    "pretrained_predictor.delete_model()\n",
    "pretrained_predictor.delete_endpoint()\n",
    "finetuned_predictor.delete_model()\n",
    "finetuned_predictor.delete_endpoint()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "759ce98f-a35a-4c64-9fae-50894b5e9f37",
   "metadata": {},
   "source": [
    "# Appendix"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7d1c8c86-bfe2-4828-a7aa-dbd7a5ee075f",
   "metadata": {},
   "source": [
    "### 1. Supported Inference Parameters\n",
    "\n",
    "---\n",
    "\n",
    "This model supports the following common payload parameters. You may specify any subset of these parameters when invoking an endpoint.\n",
    "\n",
    "- *do_sample*: If True, activates logits sampling. If specified, it must be boolean.\n",
    "- *max_new_tokens*: Maximum number of generated tokens. If specified, it must be a positive integer.\n",
    "- *repetition_penalty*: A penalty for repetitive generated text. 1.0 means no penalty.\n",
    "- *return_full_text*: If True, input text will be part of the output generated text. If specified, it must be boolean. The default value for it is False.\n",
    "- *seed*: Random sampling seed.\n",
    "- *temperature*: Controls the randomness in the output. Higher temperature results in output sequence with low-probability words and lower temperature results in output sequence with high-probability words. If `temperature` -> 0, it results in greedy decoding. If specified, it must be a positive float.\n",
    "- *top_k*: In each step of text generation, sample from only the `top_k` most likely words. If specified, it must be a positive integer.\n",
    "- *top_p*: In each step of text generation, sample from the smallest possible set of words with cumulative probability `top_p`. If specified, it must be a float between 0 and 1.\n",
    "- *details*: Return generation details, to include output token logprobs and IDs.\n",
    "\n",
    "The model will also support additional payload parameters that are dependent on the image used for this model. You can find the default image by inspecting `model.image_uri`. For information on additional payload parameters, view [LMI input output schema](https://docs.djl.ai/docs/serving/serving/docs/lmi/user_guides/lmi_input_output_schema.html) or, for text generation inference (TGI), see the following list.\n",
    "\n",
    "- *stop*: If specified, it must a list of strings. Text generation stops if any one of the specified strings is generated.\n",
    "- *truncate*: Truncate inputs tokens to the given size.\n",
    "- *typical_p*: Typical decoding mass, according to [Typical Decoding for Natural Language Generation](https://arxiv.org/abs/2202.00666).\n",
    "- *best_of*: Generate best_of sequences and return the one if the highest token logprobs.\n",
    "- *watermark*: Whether to perform watermarking with [A Watermark for Large Language Models](https://arxiv.org/abs/2301.10226).\n",
    "- *decoder_input_details*: Return decoder input token logprobs and IDs.\n",
    "- *top_n_tokens*: Return the N most likely tokens at each step.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "30d5df7e-95a5-47dc-b5d2-0178ebfc6b6f",
   "metadata": {},
   "source": [
    "### 2. Dataset formatting instruction for training\n",
    "\n",
    "---\n",
    "\n",
    "####  Fine-tune the Model on a New Dataset\n",
    "We currently offer two types of fine-tuning: instruction fine-tuning and domain adaption fine-tuning. You can easily switch to one of the training methods by specifying parameter `instruction_tuned` being 'True' or 'False'.\n",
    "\n",
    "\n",
    "#### 2.1. Domain adaptation fine-tuning\n",
    "The Text Generation model can also be fine-tuned on any domain specific dataset. After being fine-tuned on the domain specific dataset, the model is expected to generate domain specific text and solve various NLP tasks in that specific domain with **few shot prompting**.\n",
    "\n",
    "Below are the instructions for how the training data should be formatted for input to the model.\n",
    "\n",
    "- **Input:** A train and an optional validation directory. Each directory contains a CSV/JSON/TXT file. \n",
    "  - For CSV/JSON files, the train or validation data is used from the column called 'text' or the first column if no column called 'text' is found.\n",
    "  - The number of files under train and validation (if provided) should equal to one, respectively. \n",
    "- **Output:** A trained model that can be deployed for inference. \n",
    "\n",
    "Below is an example of a TXT file for fine-tuning the Text Generation model. The TXT file is SEC filings of Amazon from year 2021 to 2022.\n",
    "\n",
    "```Note About Forward-Looking Statements\n",
    "This report includes estimates, projections, statements relating to our\n",
    "business plans, objectives, and expected operating results that are “forward-\n",
    "looking statements” within the meaning of the Private Securities Litigation\n",
    "Reform Act of 1995, Section 27A of the Securities Act of 1933, and Section 21E\n",
    "of the Securities Exchange Act of 1934. Forward-looking statements may appear\n",
    "throughout this report, including the following sections: “Business” (Part I,\n",
    "Item 1 of this Form 10-K), “Risk Factors” (Part I, Item 1A of this Form 10-K),\n",
    "and “Management’s Discussion and Analysis of Financial Condition and Results\n",
    "of Operations” (Part II, Item 7 of this Form 10-K). These forward-looking\n",
    "statements generally are identified by the words “believe,” “project,”\n",
    "“expect,” “anticipate,” “estimate,” “intend,” “strategy,” “future,”\n",
    "“opportunity,” “plan,” “may,” “should,” “will,” “would,” “will be,” “will\n",
    "continue,” “will likely result,” and similar expressions. Forward-looking\n",
    "statements are based on current expectations and assumptions that are subject\n",
    "to risks and uncertainties that may cause actual results to differ materially.\n",
    "We describe risks and uncertainties that could cause actual results and events\n",
    "to differ materially in “Risk Factors,” “Management’s Discussion and Analysis\n",
    "of Financial Condition and Results of Operations,” and “Quantitative and\n",
    "Qualitative Disclosures about Market Risk” (Part II, Item 7A of this Form\n",
    "10-K). Readers are cautioned not to place undue reliance on forward-looking\n",
    "statements, which speak only as of the date they are made. We undertake no\n",
    "obligation to update or revise publicly any forward-looking statements,\n",
    "whether because of new information, future events, or otherwise.\n",
    "GENERAL\n",
    "Embracing Our Future ...\n",
    "```\n",
    "\n",
    "\n",
    "#### 2.2. Instruction fine-tuning\n",
    "\n",
    "The Text generation model can be instruction-tuned on any text data provided that the data is in the expected format. The instruction-tuned model can be further deployed for inference. Below are the instructions for how the training data should be formatted for input to the model.\n",
    "\n",
    "- **Input:** A train and an optional validation directory. Train and validation directories should contain one or multiple JSON lines (`.jsonl`) formatted files. In particular, train directory can also contain an optional `*.json` file describing the input and output formats. \n",
    "  - The best model is selected according to the validation loss, calculated at the end of each epoch.\n",
    "  If a validation set is not given, an (adjustable) percentage of the training data is\n",
    "  automatically split and used for validation.\n",
    "  - The training data must be formatted in a JSON lines (`.jsonl`) format, where each line is a dictionary\n",
    "representing a single data sample. All training data must be in a single folder, however\n",
    "it can be saved in multiple jsonl files. The `.jsonl` file extension is mandatory. The training\n",
    "folder can also contain a `template.json` file describing the input and output formats. If no\n",
    "template file is given, the following template will be used:\n",
    "  ```json\n",
    "  {\n",
    "    \"prompt\": \"Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\\n\\n### Instruction:\\n{instruction}\\n\\n### Input:\\n{context}\",\n",
    "    \"completion\": \"{response}\"\n",
    "  }\n",
    "  ```\n",
    "  - In this case, the data in the JSON lines entries must include `instruction`, `context` and `response` fields. If a custom template is provided it must also use `prompt` and `completion` keys to define\n",
    "  the input and output templates.\n",
    "  Below is a sample custom template:\n",
    "\n",
    "  ```json\n",
    "  {\n",
    "    \"prompt\": \"question: {question} context: {context}\",\n",
    "    \"completion\": \"{answer}\"\n",
    "  }\n",
    "  ```\n",
    "Here, the data in the JSON lines entries must include `question`, `context` and `answer` fields. \n",
    "- **Output:** A trained model that can be deployed for inference. \n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d4edf0c9-3c95-4e4b-932e-5d8a839d4070",
   "metadata": {},
   "source": [
    "#### 2.3. Example fine-tuning with Domain-Adaptation dataset format\n",
    "---\n",
    "We provide a subset of SEC filings data of Amazon in domain adaptation dataset format. It is downloaded from publicly available [EDGAR](https://www.sec.gov/edgar/searchedgar/companysearch). Instruction of accessing the data is shown [here](https://www.sec.gov/os/accessing-edgar-data).\n",
    "\n",
    "License: [Creative Commons Attribution-ShareAlike License (CC BY-SA 4.0)](https://creativecommons.org/licenses/by-sa/4.0/legalcode).\n",
    "\n",
    "Please uncomment the following code to fine-tune the model on dataset in domain adaptation format.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d85c93a9-ebe2-4966-a5d6-af4c053f69f2",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "import boto3\n",
    "\n",
    "model_id = \"meta-textgeneration-llama-3-2-3b\"\n",
    "\n",
    "estimator = JumpStartEstimator(\n",
    "    model_id=model_id,\n",
    "    environment={\"accept_eula\": \"true\"},\n",
    "    instance_type=\"ml.g5.12xlarge\",\n",
    ")\n",
    "estimator.set_hyperparameters(instruction_tuned=\"False\", epoch=\"5\")\n",
    "estimator.fit(\n",
    "    {\n",
    "        \"training\": f\"s3://jumpstart-cache-prod-{boto3.Session().region_name}/training-datasets/sec_amazon\"\n",
    "    }\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "19f7e4f8-970f-4a1d-b6ee-86bc77b8b9a9",
   "metadata": {},
   "source": [
    "### 3. Supported Hyper-parameters for fine-tuning\n",
    "---\n",
    "\n",
    "- **epoch** - The number of passes that the fine-tuning algorithm takes through the training dataset. Must be an integer greater than 1. Default is 5.\n",
    "- **learning_rate** - The rate at which the model weights are updated after working through each batch of training examples. Must be a positive float greater than 0. Default is 0.0001.\n",
    "- **lora_r** - Lora R dimension. Must be a positive integer. Default is 8.\n",
    "- **lora_alpha** - Lora Alpha. Must be a positive integer. Default is 32.\n",
    "- **target_modules** - Target modules for LoRA fine-tuning. You can specify a subset of [`'q_proj'`,`'v_proj'`,`'k_proj'`,`'o_proj'`,`'gate_proj'`,`'up_proj'`,`'down_proj'`] modules as a string separated by a comma without any spaces. Default is `q_proj,v_proj`.\n",
    "- **lora_dropout** - Lora Dropout. Must be a positive float between 0 and 1. Default is 0.05.\n",
    "- **instruction_tuned** - Whether to instruction-train the model or not. At most one of `instruction_tuned` and `chat_dataset` can be `True`. Must be `True` or `False`. Default is `False`.\n",
    "- **chat_dataset** - If `True`, dataset is assumed to be in chat format. At most one of `instruction_tuned` and `chat_dataset` can be `True`. Default is `False`.\n",
    "- **add_input_output_demarcation_key** - For an instruction tuned dataset, if this is `True`, a demarcation key (\"### Response:\\n\") is added between the prompt and completion before training. Default is `True`.\n",
    "- **per_device_train_batch_size** - The batch size per GPU core/CPU for training. Default is 4.\n",
    "- **per_device_eval_batch_size** - The batch size per GPU core/CPU for evaluation. Default is 1.\n",
    "- **max_train_samples** - For debugging purposes or quicker training, truncate the number of training examples to this value. Value -1 means using all of the training samples. Must be a positive integer or -1. Default is -1.\n",
    "- **max_val_samples** - For debugging purposes or quicker training, truncate the number of validation examples to this value. Value -1 means using all of the validation samples. Must be a positive integer or -1. Default is -1.\n",
    "- **seed** - Random seed that will be set at the beginning of training. Default is 10.\n",
    "- **max_input_length** - Maximum total input sequence length after tokenization. Sequences longer than this will be truncated. If -1, `max_input_length` is set to the minimum of 1024 and the maximum model length defined by the tokenizer. If set to a positive value, `max_input_length` is set to the minimum of the provided value and the `model_max_length` defined by the tokenizer. Must be a positive integer or -1. Default is -1.\n",
    "- **validation_split_ratio** - If validation channel is None, ratio of train-validation split from the train data must be between 0-1. Default is 0.2.\n",
    "- **train_data_split_seed** - If validation data is not present, this fixes the random splitting of the input training data to training and validation data used by the algorithm. Must be an integer. Default is 0.\n",
    "- **preprocessing_num_workers** - The number of processes to use for preprocessing. If None, the main process is used for preprocessing. Default is None.\n",
    "- **int8_quantization** - If `True`, the model is loaded with 8-bit precision for training. Default for Llama 3.2 1B and Llama 3.2 3B is `False`.\n",
    "- **enable_fsdp** - If `True`, training uses FSDP. Default for Llama 3.2 1B and Llama 3.2 3B is `True`.\n",
    "\n",
    "---\n",
    "\n",
    "\n",
    "- **Note 1**: int8_quantization is not supported with FSDP. Also, int8_quantization = 'False' and enable_fsdp = 'False' is not supported due to CUDA memory issues for any of the g5 family instances. Thus, we recommend setting exactly one of int8_quantization or enable_fsdp to be 'True'\n",
    "- **Note 2**: Due to the size of the model, 70B model can not be fine-tuned with enable_fsdp = 'True' for any of the supported instance types.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "72b6d023-3487-4571-8b52-f332790c1ad7",
   "metadata": {},
   "source": [
    "### 4. Supported Instance types for fine-tuning Llama 3.2\n",
    "\n",
    "---\n",
    "We have tested our scripts on the following instances types for fine-tuning Llama 3:\n",
    "\n",
    "| Model | Model ID | All Supported Instances Types for fine-tuning |\n",
    "| - | - | - |\n",
    "| Llama 3.2 1B | meta-textgeneration-llama-3-2-1b | ml.g5.2xlarge, ml.g5.4xlarge, ml.g5.8xlarge, ml.g5.12xlarge, ml.p3dn.24xlarge, ml.g4dn.12xlarge, ml.p5.48xlarge |\n",
    "| Llama 3.2 1B Instruct | meta-textgeneration-llama-3-2-1b-instruct | ml.g5.2xlarge, ml.g5.4xlarge, ml.g5.8xlarge, ml.g5.12xlarge, ml.p3dn.24xlarge, ml.g4dn.12xlarge, ml.p5.48xlarge |\n",
    "| Llama 3.2 3B | meta-textgeneration-llama-3-2-3b | ml.g5.12xlarge, ml.g5.24xlarge, ml.g5.48xlarge, ml.p3dn.24xlarge, ml.g4dn.12xlarge, ml.p5.48xlarge |\n",
    "| Llama 3.2 3B Instruct | meta-textgeneration-llama-3-2-3b-instruct | ml.g5.12xlarge, ml.g5.24xlarge, ml.g5.48xlarge, ml.p3dn.24xlarge, ml.g4dn.12xlarge, ml.p5.48xlarge |\n",
    "\n",
    "Other instance types may also work to fine-tune. Note: When using p3 instances, training will be done with 32 bit precision as bfloat16 is not supported on these instances. Thus, training job would consume double the amount of CUDA memory when training on p3 instances compared to g5 instances.\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "770d4350-cf4d-40a0-be1c-eba44efd33ab",
   "metadata": {},
   "source": [
    "### 5. Few notes about the fine-tuning method\n",
    "\n",
    "---\n",
    "- Fine-tuning scripts are based on [this repo](https://github.com/facebookresearch/llama-recipes/tree/main). \n",
    "- Instruction tuning dataset is first converted into domain adaptation dataset format before fine-tuning. \n",
    "- Fine-tuning scripts utilize Fully Sharded Data Parallel (FSDP) as well as Low Rank Adaptation (LoRA) method fine-tuning the models\n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "909ce841-3a2c-4c08-a102-b94148036a5a",
   "metadata": {},
   "source": [
    "### 6. Studio Kernel Dead/Creating JumpStart Model from the training Job\n",
    "---\n",
    "Due to the size of the Llama 70B model, training job may take several hours and the studio kernel may die during the training phase. However, during this time, training is still running in SageMaker. If this happens, you can still deploy the endpoint using the training job name with the following code:\n",
    "\n",
    "How to find the training job name? Go to Console -> SageMaker -> Training -> Training Jobs -> Identify the training job name and substitute in the following cell. \n",
    "\n",
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "dfa60a66-1c2f-42df-8079-191319e28a65",
   "metadata": {
    "scrolled": true,
    "tags": []
   },
   "outputs": [],
   "source": [
    "from sagemaker.jumpstart.estimator import JumpStartEstimator\n",
    "\n",
    "training_job_name = \"<<Replace this with Training Job Name>>\"\n",
    "\n",
    "attached_estimator = JumpStartEstimator.attach(training_job_name, model_id)\n",
    "attached_estimator.logs()\n",
    "attached_estimator.deploy()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a7be1219",
   "metadata": {},
   "source": [
    "## Notebook CI Test Results\n",
    "\n",
    "This notebook was tested in multiple regions. The test results are as follows, except for us-west-2 which is shown at the top of the notebook.\n",
    "\n",
    "\n",
    "![This us-east-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/us-east-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This us-east-2 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/us-east-2/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This us-west-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/us-west-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This ca-central-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/ca-central-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This sa-east-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/sa-east-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This eu-west-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/eu-west-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This eu-west-2 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/eu-west-2/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This eu-west-3 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/eu-west-3/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This eu-central-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/eu-central-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This eu-north-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/eu-north-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This ap-southeast-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/ap-southeast-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This ap-southeast-2 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/ap-southeast-2/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This ap-northeast-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/ap-northeast-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This ap-northeast-2 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/ap-northeast-2/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n",
    "\n",
    "![This ap-south-1 badge failed to load. Check your device's internet connectivity, otherwise the service is currently unavailable](https://prod.us-west-2.tcx-beacon.docs.aws.dev/sagemaker-nb/ap-south-1/generative_ai|sm-jumpstart_foundation_llama_3_2_3b_finetuning.ipynb)\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.14"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}