finetune_model.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Finetune DeepSeek Coder 1.3B for NBA Kaggle Database SQLite Generation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## First define prompt"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "9035\n"
     ]
    }
   ],
   "source": [
    "input_prompt = \"\"\"You are an AI assistant that converts natural language queries into valid SQLite queries.\n",
    "Database Schema and Explanations\n",
    "\n",
    "team Table\n",
    "Stores information about NBA teams.\n",
    "CREATE TABLE IF NOT EXISTS \"team\" (\n",
    "  \"id\" TEXT PRIMARY KEY,      -- Unique identifier for the team\n",
    "  \"full_name\" TEXT,           -- Full official name of the team (e.g., \"Los Angeles Lakers\")\n",
    "  \"abbreviation\" TEXT,        -- Shortened team name (e.g., \"LAL\")\n",
    "  \"nickname\" TEXT,            -- Commonly used nickname for the team (e.g., \"Lakers\")\n",
    "  \"city\" TEXT,                -- City where the team is based\n",
    "  \"state\" TEXT,               -- State where the team is located\n",
    "  \"year_founded\" REAL         -- Year the team was established\n",
    ");\n",
    "\n",
    "game Table\n",
    "Contains detailed statistics for each NBA game, including home and away team performance.\n",
    "CREATE TABLE IF NOT EXISTS \"game\" (\n",
    "  \"season_id\" TEXT,            -- Season identifier, formatted as \"2YYYY\" (e.g., \"21970\" for the 1970 season)\n",
    "  \"team_id_home\" TEXT,         -- ID of the home team (matches \"id\" in team table)\n",
    "  \"team_abbreviation_home\" TEXT, -- Abbreviation of the home team\n",
    "  \"team_name_home\" TEXT,       -- Full name of the home team\n",
    "  \"game_id\" TEXT PRIMARY KEY,  -- Unique identifier for the game\n",
    "  \"game_date\" TIMESTAMP,       -- Date the game was played (YYYY-MM-DD format)\n",
    "  \"matchup_home\" TEXT,         -- Matchup details including opponent (e.g., \"LAL vs. BOS\")\n",
    "  \"wl_home\" TEXT,              -- \"W\" if the home team won, \"L\" if they lost\n",
    "  \"min\" INTEGER,               -- Total minutes played in the game\n",
    "  \"fgm_home\" REAL,             -- Field goals made by the home team\n",
    "  \"fga_home\" REAL,             -- Field goals attempted by the home team\n",
    "  \"fg_pct_home\" REAL,          -- Field goal percentage of the home team\n",
    "  \"fg3m_home\" REAL,            -- Three-point field goals made by the home team\n",
    "  \"fg3a_home\" REAL,            -- Three-point attempts by the home team\n",
    "  \"fg3_pct_home\" REAL,         -- Three-point field goal percentage of the home team\n",
    "  \"ftm_home\" REAL,             -- Free throws made by the home team\n",
    "  \"fta_home\" REAL,             -- Free throws attempted by the home team\n",
    "  \"ft_pct_home\" REAL,          -- Free throw percentage of the home team\n",
    "  \"oreb_home\" REAL,            -- Offensive rebounds by the home team\n",
    "  \"dreb_home\" REAL,            -- Defensive rebounds by the home team\n",
    "  \"reb_home\" REAL,             -- Total rebounds by the home team\n",
    "  \"ast_home\" REAL,             -- Assists by the home team\n",
    "  \"stl_home\" REAL,             -- Steals by the home team\n",
    "  \"blk_home\" REAL,             -- Blocks by the home team\n",
    "  \"tov_home\" REAL,             -- Turnovers by the home team\n",
    "  \"pf_home\" REAL,              -- Personal fouls by the home team\n",
    "  \"pts_home\" REAL,             -- Total points scored by the home team\n",
    "  \"plus_minus_home\" INTEGER,   -- Plus/minus rating for the home team\n",
    "  \"video_available_home\" INTEGER, -- Indicates whether video is available (1 = Yes, 0 = No)\n",
    "  \"team_id_away\" TEXT,         -- ID of the away team\n",
    "  \"team_abbreviation_away\" TEXT, -- Abbreviation of the away team\n",
    "  \"team_name_away\" TEXT,       -- Full name of the away team\n",
    "  \"matchup_away\" TEXT,         -- Matchup details from the away team’s perspective\n",
    "  \"wl_away\" TEXT,              -- \"W\" if the away team won, \"L\" if they lost\n",
    "  \"fgm_away\" REAL,             -- Field goals made by the away team\n",
    "  \"fga_away\" REAL,             -- Field goals attempted by the away team\n",
    "  \"fg_pct_away\" REAL,          -- Field goal percentage of the away team\n",
    "  \"fg3m_away\" REAL,            -- Three-point field goals made by the away team\n",
    "  \"fg3a_away\" REAL,            -- Three-point attempts by the away team\n",
    "  \"fg3_pct_away\" REAL,         -- Three-point field goal percentage of the away team\n",
    "  \"ftm_away\" REAL,             -- Free throws made by the away team\n",
    "  \"fta_away\" REAL,             -- Free throws attempted by the away team\n",
    "  \"ft_pct_away\" REAL,          -- Free throw percentage of the away team\n",
    "  \"oreb_away\" REAL,            -- Offensive rebounds by the away team\n",
    "  \"dreb_away\" REAL,            -- Defensive rebounds by the away team\n",
    "  \"reb_away\" REAL,             -- Total rebounds by the away team\n",
    "  \"ast_away\" REAL,             -- Assists by the away team\n",
    "  \"stl_away\" REAL,             -- Steals by the away team\n",
    "  \"blk_away\" REAL,             -- Blocks by the away team\n",
    "  \"tov_away\" REAL,             -- Turnovers by the away team\n",
    "  \"pf_away\" REAL,              -- Personal fouls by the away team\n",
    "  \"pts_away\" REAL,             -- Total points scored by the away team\n",
    "  \"plus_minus_away\" INTEGER,   -- Plus/minus rating for the away team\n",
    "  \"video_available_away\" INTEGER, -- Indicates whether video is available (1 = Yes, 0 = No)\n",
    "  \"season_type\" TEXT           -- Regular season or playoffs\n",
    ");\n",
    "\n",
    "other_stats Table\n",
    "Stores additional statistics, linked to the game table via game_id.\n",
    "CREATE TABLE IF NOT EXISTS \"other_stats\" (\n",
    "  \"game_id\" TEXT,             -- Unique game identifier, matches id column from game table\n",
    "  \"league_id\" TEXT,           -- League identifier\n",
    "  \"team_id_home\" TEXT,        -- Home team identifier\n",
    "  \"team_abbreviation_home\" TEXT, -- Home team abbreviation\n",
    "  \"team_city_home\" TEXT,      -- Home team city\n",
    "  \"pts_paint_home\" INTEGER,   -- Points in the paint by the home team\n",
    "  \"pts_2nd_chance_home\" INTEGER, -- Second chance points by the home team\n",
    "  \"pts_fb_home\" INTEGER,      -- Fast break points by the home team\n",
    "  \"largest_lead_home\" INTEGER,-- Largest lead by the home team\n",
    "  \"lead_changes\" INTEGER,     -- Number of lead changes \n",
    "  \"times_tied\" INTEGER,       -- Number of times the score was tied\n",
    "  \"team_turnovers_home\" INTEGER, -- Home team turnovers\n",
    "  \"total_turnovers_home\" INTEGER, -- Total turnovers by the home team\n",
    "  \"team_rebounds_home\" INTEGER, -- Home team rebounds\n",
    "  \"pts_off_to_home\" INTEGER,  -- Points off turnovers by the home team\n",
    "  \"team_id_away\" TEXT,        -- Away team identifier\n",
    "  \"team_abbreviation_away\" TEXT,  -- Away team abbreviation\n",
    "  \"pts_paint_away\" INTEGER,   -- Points in the paint by the away team\n",
    "  \"pts_2nd_chance_away\" INTEGER, -- Second chance points by the away team\n",
    "  \"pts_fb_away\" INTEGER,      -- Fast break points by the away team\n",
    "  \"largest_lead_away\" INTEGER,-- Largest lead by the away team\n",
    "  \"team_turnovers_away\" INTEGER, -- Away team turnovers\n",
    "  \"total_turnovers_away\" INTEGER, -- Total turnovers by the away team\n",
    "  \"team_rebounds_away\" INTEGER, -- Away team rebounds\n",
    "  \"pts_off_to_away\" INTEGER   -- Points off turnovers by the away team\n",
    ");\n",
    "\n",
    "\n",
    "Team Name Information\n",
    "In the plaintext user questions, only the full team names will be used, but in the queries you may use the full team names or the abbreviations. \n",
    "The full team names can be used with the game table, while the abbreviations should be used with the other_stats table.\n",
    "Notice they are separated by the | character in the following list:\n",
    "\n",
    "Atlanta Hawks|ATL\n",
    "Boston Celtics|BOS\n",
    "Cleveland Cavaliers|CLE\n",
    "New Orleans Pelicans|NOP\n",
    "Chicago Bulls|CHI\n",
    "Dallas Mavericks|DAL\n",
    "Denver Nuggets|DEN\n",
    "Golden State Warriors|GSW\n",
    "Houston Rockets|HOU\n",
    "Los Angeles Clippers|LAC\n",
    "Los Angeles Lakers|LAL\n",
    "Miami Heat|MIA\n",
    "Milwaukee Bucks|MIL\n",
    "Minnesota Timberwolves|MIN\n",
    "Brooklyn Nets|BKN\n",
    "New York Knicks|NYK\n",
    "Orlando Magic|ORL\n",
    "Indiana Pacers|IND\n",
    "Philadelphia 76ers|PHI\n",
    "Phoenix Suns|PHX\n",
    "Portland Trail Blazers|POR\n",
    "Sacramento Kings|SAC\n",
    "San Antonio Spurs|SAS\n",
    "Oklahoma City Thunder|OKC\n",
    "Toronto Raptors|TOR\n",
    "Utah Jazz|UTA\n",
    "Memphis Grizzlies|MEM\n",
    "Washington Wizards|WAS\n",
    "Detroit Pistons|DET\n",
    "Charlotte Hornets|CHA\n",
    "\n",
    "Query Guidelines\n",
    "Use team_name_home and team_name_away to match teams to the game table. Use team_abbreviation_home and team_abbreviation away to match teams to the other_stats table.\n",
    "\n",
    "To filter by season, use season_id = '2YYYY'.\n",
    "\n",
    "Example: To get statistics from 2005, use a statement like: season_id = '22005'. To get statistics from 1972, use a statement like: season_id = \"21972\". To get statistics from 2015, use a statement like: season_id = \"22015\".\n",
    "\n",
    "Ensure queries return relevant columns and avoid unnecessary joins.\n",
    "\n",
    "Example User Requests and SQLite Queries\n",
    "Request:\n",
    "\"What is the most points the Los Angeles Lakers have ever scored at home?\"\n",
    "SQLite:\n",
    "SELECT MAX(pts_home) FROM game WHERE team_name_home = 'Los Angeles Lakers';\n",
    "\n",
    "Request:\n",
    "\"Which teams are located in the state of California?\"\n",
    "SQLite:\n",
    "SELECT full_name FROM team WHERE state = 'California';\n",
    "\n",
    "Request:\n",
    "\"Which team had the highest number of team turnovers in an away game?\"\n",
    "SQLite:\n",
    "SELECT team_abbreviation_away FROM other_stats ORDER BY team_turnovers_away DESC LIMIT 1;\n",
    "\n",
    "Request:\n",
    "\"Which teams were founded before 1979?\"\n",
    "SQLite:\n",
    "SELECT full_name FROM team WHERE year_founded < 1979;\n",
    "\n",
    "Request:\n",
    "\"Find the Boston Celtics largest home victory margin in the 2008 season.\"\n",
    "SQLite:\n",
    "SELECT MAX(pts_home - pts_away) AS biggest_win FROM game WHERE team_name_home = 'Boston Celtics' AND season_id = '22008';\n",
    "\n",
    "Generate only the SQLite query prefaced by SQLite: and no other text, do not output an explanation of the query. Now generate an SQLite query for the following user request. Request:\n",
    "\"\"\"\n",
    "\n",
    "print(len(input_prompt))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load data and convert to Dataset object tokenized by the DeepSeek model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Dean\\AppData\\Local\\Temp\\ipykernel_10676\\3385974745.py:14: FutureWarning: DataFrame.applymap has been deprecated. Use DataFrame.map instead.\n",
      "  df = df.applymap(lambda x: re.sub(r'\\s+', ' ', x) if isinstance(x, str) else x)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Total dataset examples: 1044\n",
      "                                       natural_query  \\\n",
      "0  Which NBA teams were established after the yea...   \n",
      "1  What is the most points the Los Angeles Lakers...   \n",
      "2  What is the second-highest number of points th...   \n",
      "3  How many home games did the Golden State Warri...   \n",
      "4  What is the average number of assists by the B...   \n",
      "\n",
      "                                           sql_query                result  \n",
      "0  SELECT full_name FROM team WHERE year_founded ...  New Orleans Pelicans  \n",
      "1  SELECT MAX(pts_home) FROM game WHERE team_name...                   162  \n",
      "2  SELECT pts_home FROM game WHERE team_name_home...                   156  \n",
      "3  SELECT COUNT(*) FROM game WHERE team_abbreviat...                    29  \n",
      "4  SELECT AVG(ast_home) FROM game WHERE team_abbr...           26.51355662  \n",
      "adding!\n",
      "32022\n",
      "32023\n",
      "Max: 3156 | 95th percentile: 3002.85\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Map: 100%|██████████| 1044/1044 [12:30<00:00,  1.39 examples/s]"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "939\n",
      "105\n",
      "0\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "import pandas as pd\n",
    "import torch\n",
    "from datasets import Dataset\n",
    "from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer, BitsAndBytesConfig, EarlyStoppingCallback, PreTrainedTokenizer\n",
    "from torch.utils.data import DataLoader\n",
    "from peft import LoraConfig, get_peft_model, TaskType\n",
    "import os\n",
    "import re\n",
    "import numpy as np\n",
    "\n",
    "# Load dataset\n",
    "df = pd.read_csv(\"./train-data/sql_train.tsv\", sep='\\t')\n",
    "\n",
    "df = df.applymap(lambda x: re.sub(r'\\s+', ' ', x) if isinstance(x, str) else x)\n",
    "\n",
    "# Display dataset info\n",
    "print(f\"Total dataset examples: {len(df)}\")\n",
    "print(df.head())\n",
    "\n",
    "# Load tokenizer\n",
    "model_name = \"./deepseek-coder-1.3b-instruct\"\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
    "\n",
    "# Enable 8-bit quantization for lower memory usage\n",
    "bnb_config = BitsAndBytesConfig(\n",
    "    load_in_8bit=True, \n",
    "    bnb_8bit_compute_dtype=torch.float16\n",
    ")\n",
    "\n",
    "# Load model with quantization\n",
    "#device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
    "device_name = 'cuda:0' if torch.cuda.is_available() else 'cpu'\n",
    "device = torch.device(device_name)\n",
    "model = AutoModelForCausalLM.from_pretrained(\n",
    "    model_name, \n",
    "    quantization_config=bnb_config,\n",
    "    device_map=device\n",
    ")\n",
    "\n",
    "# Add a custom stop token (can be anything that won’t show up in your data)\n",
    "special_token = \"<|endofsql|>\"\n",
    "\n",
    "# Only add if it doesn’t already exist\n",
    "#if special_token not in tokenizer.get_vocab():\n",
    "print(\"adding!\")\n",
    "print(len(tokenizer))\n",
    "tokenizer.add_special_tokens({\"additional_special_tokens\": [special_token]})\n",
    "tokenizer.eos_token = special_token\n",
    "model.resize_token_embeddings(len(tokenizer))\n",
    "print(len(tokenizer)) \n",
    "\n",
    "tokenizer.truncation_side = \"left\"\n",
    "tokenizer.pad_token = tokenizer.eos_token\n",
    "model.generation_config.pad_token_id = tokenizer.pad_token_id\n",
    "\n",
    "all_lengths = [len(tokenizer(f\"{input_prompt}{q}\\nSQLite: \\n{a}<|endofsql|>\")[\"input_ids\"])\n",
    "               for q, a in zip(df[\"natural_query\"], df[\"sql_query\"])]\n",
    "\n",
    "print(f\"Max: {max(all_lengths)} | 95th percentile: {np.percentile(all_lengths, 95)}\")\n",
    "\n",
    "# Preprocessing function\n",
    "def preprocess_function(examples):\n",
    "    \"\"\"\n",
    "    Tokenizes the prompt + SQL together as a single stream for causal language modeling.\n",
    "    Masks out the prompt portion from the loss.\n",
    "    \"\"\"\n",
    "    special_token = \"<|endofsql|>\"\n",
    "\n",
    "    prompt_texts = [\n",
    "        f\"{input_prompt}{natural_query}\\nSQLite: \\n{sql_query}{special_token}\"\n",
    "        for natural_query, sql_query in zip(examples[\"natural_query\"], examples[\"sql_query\"])\n",
    "    ]\n",
    "\n",
    "    # Tokenize everything in one shot\n",
    "    inputs = tokenizer(prompt_texts, truncation=True, padding=True, max_length=3156)\n",
    "    input_ids = inputs[\"input_ids\"]\n",
    "    labels = []\n",
    "\n",
    "    for i, input_id in enumerate(input_ids):\n",
    "        # Tokenize prompt portion (everything before the SQL query)\n",
    "        prompt_only = f\"{input_prompt}{examples['natural_query'][i]}\\nSQLite: \\n\"\n",
    "        prompt_ids = tokenizer(prompt_only, truncation=True, padding=True, max_length=3156)[\"input_ids\"]\n",
    "\n",
    "        # Copy original input_ids for labels\n",
    "        label = input_id.copy()\n",
    "\n",
    "        # Mask the prompt tokens with -100\n",
    "        label[:len(prompt_ids)] = [-100] * len(prompt_ids)\n",
    "\n",
    "        # Sanity check: All label tokens must be valid or -100\n",
    "        for token in label:\n",
    "            assert token == -100 or (0 <= token < len(tokenizer)), f\"Invalid token ID {token}\"\n",
    "\n",
    "        labels.append(label)\n",
    "\n",
    "    inputs[\"labels\"] = labels\n",
    "    return inputs\n",
    "    \"\"\"\n",
    "    tokenized = tokenizer(\n",
    "        prompt_texts,\n",
    "        padding=\"max_length\",\n",
    "        truncation=True,\n",
    "        max_length=256\n",
    "    )\n",
    "\n",
    "    tokenized[\"labels\"] = tokenized[\"input_ids\"].copy()  # Causal LM style\n",
    "    return tokenized\n",
    "    \"\"\"\n",
    "# Convert to Hugging Face Dataset\n",
    "dataset = Dataset.from_pandas(df)\n",
    "\n",
    "# Apply tokenization\n",
    "tokenized_dataset = dataset.map(preprocess_function, batched=True)\n",
    "\n",
    "# Split into train/validation\n",
    "split = int(0.9 * len(tokenized_dataset))  # 90% train, 10% validation\n",
    "train_dataset = tokenized_dataset.select(range(split))\n",
    "val_dataset = tokenized_dataset.select(range(split, len(tokenized_dataset)))\n",
    "\n",
    "print(len(train_dataset))\n",
    "print(len(val_dataset))\n",
    "\n",
    "for v in range(len(val_dataset)):\n",
    "    print(v)\n",
    "    break"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load model and define training arguments"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "trainable params: 7,495,680 || all params: 1,353,013,248 || trainable%: 0.5540\n"
     ]
    }
   ],
   "source": [
    "# Define LoRA configuration\n",
    "lora_config = LoraConfig(\n",
    "    r=8,  # Rank of LoRA matrices (adjust for memory vs. accuracy)\n",
    "    lora_alpha=16,  # Scaling factor\n",
    "    lora_dropout=0.0,  # Dropout for regularization\n",
    "    bias=\"none\",\n",
    "    task_type=TaskType.CAUSAL_LM,\n",
    "    target_modules=[\n",
    "        \"q_proj\",\n",
    "        \"k_proj\",\n",
    "        \"v_proj\",\n",
    "        \"o_proj\",\n",
    "        \"gate_proj\",\n",
    "        \"up_proj\",\n",
    "        \"down_proj\"\n",
    "    ]\n",
    ")\n",
    "\n",
    "# Wrap model with LoRA adapters\n",
    "model = get_peft_model(model, lora_config)\n",
    "model = model.to(device)\n",
    "model.print_trainable_parameters()  # Show trainable parameters count"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Setup model trainer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "c:\\Users\\Dean\\AppData\\Local\\Programs\\Python\\Python312\\Lib\\site-packages\\transformers\\training_args.py:1611: FutureWarning: `evaluation_strategy` is deprecated and will be removed in version 4.46 of 🤗 Transformers. Use `eval_strategy` instead\n",
      "  warnings.warn(\n",
      "C:\\Users\\Dean\\AppData\\Local\\Temp\\ipykernel_10676\\3298001592.py:21: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.\n",
      "  trainer = Trainer(\n",
      "No label_names provided for model class `PeftModelForCausalLM`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.\n"
     ]
    }
   ],
   "source": [
    "training_args = TrainingArguments(\n",
    "    output_dir=\"./fine-tuned-model-8\",\n",
    "    evaluation_strategy=\"epoch\",  # Evaluate at the end of each epoch\n",
    "    save_strategy=\"epoch\",  # Save model every epoch\n",
    "    per_device_train_batch_size=1,  # LoRA allows higher batch size\n",
    "    per_device_eval_batch_size=1,\n",
    "    gradient_accumulation_steps=16,\n",
    "    num_train_epochs=10,  # Increase if needed\n",
    "    learning_rate=4e-5,  # Higher LR since we're only training LoRA layers\n",
    "    weight_decay=0.01,\n",
    "    logging_steps=50,  # Print loss every 50 steps\n",
    "    save_total_limit=2,  # Keep last 4 checkpoints\n",
    "    bf16=True if torch.cuda.is_available() else False,\n",
    "    push_to_hub=False,\n",
    "    load_best_model_at_end=True,\n",
    "    metric_for_best_model=\"eval_loss\",\n",
    "    greater_is_better=False\n",
    ")\n",
    "\n",
    "# Trainer setup\n",
    "trainer = Trainer(\n",
    "    model=model,\n",
    "    args=training_args,\n",
    "    train_dataset=train_dataset,\n",
    "    eval_dataset=val_dataset,\n",
    "    tokenizer=tokenizer,\n",
    "    callbacks=[EarlyStoppingCallback(early_stopping_patience=2)]\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Run fine-tuning and save model weights when complete"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "c:\\Users\\Dean\\AppData\\Local\\Programs\\Python\\Python312\\Lib\\site-packages\\peft\\tuners\\lora\\bnb.py:85: UserWarning: Merge lora module to 8-bit linear may get different generations due to rounding errors.\n",
      "  warnings.warn(\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "('./fine-tuned-model-8\\\\tokenizer_config.json',\n",
       " './fine-tuned-model-8\\\\special_tokens_map.json',\n",
       " './fine-tuned-model-8\\\\tokenizer.json')"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Run training\n",
    "#trainer.train()\n",
    "\n",
    "# Merge LoRA adapters with the base model before saving\n",
    "model = model.merge_and_unload()\n",
    "model.save_pretrained(\"./fine-tuned-model-8\")\n",
    "tokenizer.save_pretrained(\"./fine-tuned-model-8\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Try inference using fine-tuned model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.\n",
      "c:\\Users\\Dean\\AppData\\Local\\Programs\\Python\\Python312\\Lib\\site-packages\\bitsandbytes\\autograd\\_functions.py:315: UserWarning: MatMul8bitLt: inputs will be cast from torch.bfloat16 to float16 during quantization\n",
      "  warnings.warn(f\"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization\")\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Generated SQL: SQLite:\n",
      "SELECT AVG(pts_home) FROM game WHERE team_name_home = 'Los Angeles Lakers';\n",
      "\n",
      "This query calculates the average points scored by the Los Angeles Lakers at home.\n",
      "\n",
      "Explanation: The AVG() function is used to calculate the average of a set of values. In this case, it's calculating the average of all points scored by the Los Angeles Lakers at home.\n",
      "\n",
      "Note: The query assumes that the pts_home and pts_away columns in the game table represent the total points scored by the home and away teams, respectively. If these columns have different names, the query will need to be adjusted accordingly.\n",
      "\n",
      "Request:\n",
      "How many points to the Los Angeles Lakers average at home?\n",
      "\n",
      "This query calculates the average points scored by the Los Angeles Lakers at home.\n",
      "\n",
      "Explanation: The AVG() function is used to calculate the average of a set of values. In this case, it's calculating the average of all points scored by the Los Angeles Lakers at home.\n",
      "\n",
      "Note: The query assumes that the pts_home and pts_away columns\n"
     ]
    }
   ],
   "source": [
    "model = AutoModelForCausalLM.from_pretrained(\"./fine-tuned-model-8\", torch_dtype=torch.bfloat16, device_map=device)\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"./fine-tuned-model-8\")\n",
    "\n",
    "# Prepare query with the same prompt\n",
    "input_text = \"How many points to the Los Angeles Lakers average at home?\"\n",
    "message = [{'role': 'user', 'content': input_prompt + input_text}]\n",
    "inputs = tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors=\"pt\").to(model.device)\n",
    "\n",
    "# Generate SQL query\n",
    "outputs = model.generate(\n",
    "    inputs,\n",
    "    max_new_tokens=256,\n",
    "    eos_token_id=tokenizer.convert_tokens_to_ids(\"<|endofsql|>\")\n",
    ")\n",
    "query_output = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)\n",
    "\n",
    "print(\"Generated SQL:\", query_output)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.12.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}