03_Instruction_tuning.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "501de373",
   "metadata": {},
   "source": [
    "# Instruction-tuning"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5729f6ce",
   "metadata": {},
   "outputs": [],
   "source": [
    "import itertools\n",
    "import jsonlines\n",
    "\n",
    "from datasets import load_dataset\n",
    "from pprint import pprint\n",
    "\n",
    "from llama import BasicModelRunner\n",
    "from transformers import AutoTokenizer, AutoModelForCausalLM\n",
    "from transformers import AutoModelForSeq2SeqLM, AutoTokenizer"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4a712487",
   "metadata": {},
   "source": [
    "## Load instruction tuned dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1a4ed11a",
   "metadata": {},
   "outputs": [],
   "source": [
    "instruction_tuned_dataset = load_dataset(\"tatsu-lab/alpaca\", split=\"train\", streaming=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9e48415c",
   "metadata": {},
   "outputs": [],
   "source": [
    "m = 5\n",
    "print(\"Instruction-tuned dataset:\")\n",
    "top_m = list(itertools.islice(instruction_tuned_dataset, m))\n",
    "for j in top_m:\n",
    "  print(j)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "54f2cf23",
   "metadata": {},
   "source": [
    "## Two prompt templates"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "441aeced",
   "metadata": {},
   "outputs": [],
   "source": [
    "prompt_template_with_input = \"\"\"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",
    "{input}\n",
    "\n",
    "### Response:\"\"\"\n",
    "\n",
    "prompt_template_without_input = \"\"\"Below is an instruction that describes a task. Write a response that appropriately completes the request.\n",
    "\n",
    "### Instruction:\n",
    "{instruction}\n",
    "\n",
    "### Response:\"\"\""
   ]
  },
  {
   "cell_type": "markdown",
   "id": "26a977e7",
   "metadata": {},
   "source": [
    "## Hydrate prompts (add data to prompts)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "70506882",
   "metadata": {},
   "outputs": [],
   "source": [
    "processed_data = []\n",
    "for j in top_m:\n",
    "  if not j[\"input\"]:\n",
    "    processed_prompt = prompt_template_without_input.format(instruction=j[\"instruction\"])\n",
    "  else:\n",
    "    processed_prompt = prompt_template_with_input.format(instruction=j[\"instruction\"], input=j[\"input\"])\n",
    "\n",
    "  processed_data.append({\"input\": processed_prompt, \"output\": j[\"output\"]})\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3f176a0b",
   "metadata": {},
   "outputs": [],
   "source": [
    "pprint(processed_data[0])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6ad87987",
   "metadata": {},
   "source": [
    "## Save data to jsonl"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d8186af6",
   "metadata": {},
   "outputs": [],
   "source": [
    "with jsonlines.open(f'alpaca_processed.jsonl', 'w') as writer:\n",
    "    writer.write_all(processed_data)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1595cc48",
   "metadata": {},
   "source": [
    "## Compare non-instruction-tuned vs. instruction-tuned models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "03e82d02",
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset_path_hf = \"lamini/alpaca\"\n",
    "dataset_hf = load_dataset(dataset_path_hf)\n",
    "print(dataset_hf)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c03cb66d",
   "metadata": {},
   "outputs": [],
   "source": [
    "non_instruct_model = BasicModelRunner(\"meta-llama/Llama-2-7b-hf\")\n",
    "non_instruct_output = non_instruct_model(\"Tell me how to train my dog to sit\")\n",
    "print(\"Not instruction-tuned output (Llama 2 Base):\", non_instruct_output)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c04c66bf",
   "metadata": {},
   "outputs": [],
   "source": [
    "instruct_model = BasicModelRunner(\"meta-llama/Llama-2-7b-chat-hf\")\n",
    "instruct_output = instruct_model(\"Tell me how to train my dog to sit\")\n",
    "print(\"Instruction-tuned output (Llama 2): \", instruct_output)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a45f3840",
   "metadata": {},
   "outputs": [],
   "source": [
    "chatgpt = BasicModelRunner(\"chat-gpt\")\n",
    "instruct_output_chatgpt = chatgpt(\"Tell me how to train my dog to sit\")\n",
    "print(\"Instruction-tuned output (ChatGPT): \", instruct_output_chatgpt)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "277b4ad5",
   "metadata": {},
   "source": [
    "## Try smaller models"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c37ef86e",
   "metadata": {},
   "outputs": [],
   "source": [
    "tokenizer = AutoTokenizer.from_pretrained(\"EleutherAI/pythia-70m\")\n",
    "model = AutoModelForCausalLM.from_pretrained(\"EleutherAI/pythia-70m\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "afe5ede0",
   "metadata": {},
   "outputs": [],
   "source": [
    "def inference(text, model, tokenizer, max_input_tokens=1000, max_output_tokens=100):\n",
    "  # Tokenize\n",
    "  input_ids = tokenizer.encode(\n",
    "          text,\n",
    "          return_tensors=\"pt\",\n",
    "          truncation=True,\n",
    "          max_length=max_input_tokens\n",
    "  )\n",
    "\n",
    "  # Generate\n",
    "  device = model.device\n",
    "  generated_tokens_with_prompt = model.generate(\n",
    "    input_ids=input_ids.to(device),\n",
    "    max_length=max_output_tokens\n",
    "  )\n",
    "\n",
    "  # Decode\n",
    "  generated_text_with_prompt = tokenizer.batch_decode(generated_tokens_with_prompt, skip_special_tokens=True)\n",
    "\n",
    "  # Strip the prompt\n",
    "  generated_text_answer = generated_text_with_prompt[0][len(text):]\n",
    "\n",
    "  return generated_text_answer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "22f5e568",
   "metadata": {},
   "outputs": [],
   "source": [
    "finetuning_dataset_path = \"lamini/lamini_docs\"\n",
    "finetuning_dataset = load_dataset(finetuning_dataset_path)\n",
    "print(finetuning_dataset)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b77abca6",
   "metadata": {},
   "outputs": [],
   "source": [
    "test_sample = finetuning_dataset[\"test\"][0]\n",
    "print(test_sample)\n",
    "\n",
    "print(inference(test_sample[\"question\"], model, tokenizer))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8bfe428d",
   "metadata": {},
   "source": [
    "## Compare to finetuned small model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e3df65ed",
   "metadata": {},
   "outputs": [],
   "source": [
    "instruction_model = AutoModelForCausalLM.from_pretrained(\"lamini/lamini_docs_finetuned\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9e4c39c3",
   "metadata": {},
   "outputs": [],
   "source": [
    "print(inference(test_sample[\"question\"], instruction_model, tokenizer))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3da151c2",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Pssst! If you were curious how to upload your own dataset to Huggingface\n",
    "# Here is how we did it\n",
    "\n",
    "# !pip install huggingface_hub\n",
    "# !huggingface-cli login\n",
    "\n",
    "# import pandas as pd\n",
    "# import datasets\n",
    "# from datasets import Dataset\n",
    "\n",
    "# finetuning_dataset = Dataset.from_pandas(pd.DataFrame(data=finetuning_dataset))\n",
    "# finetuning_dataset.push_to_hub(dataset_path_hf)"
   ]
  }
 ],
 "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": "3.9.16"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}