{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "# Run pre-trained DeepSeek Coder 1.3B Model on Chat-GPT 4o generated dataset" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## First load dataset into pandas dataframe" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Total dataset examples: 1044\n", "\n", "\n", "What was the largest lead the Golden State Warriors had in a game during the 2018 season?\n", "SELECT MAX(other_stats.largest_lead_home) FROM other_stats JOIN game ON other_stats.game_id = game.game_id WHERE game.team_name_home = 'Golden State Warriors' AND game.season_id = '22018';\n", "44\n" ] } ], "source": [ "import pandas as pd \n", "\n", "# Load dataset and check length\n", "df = pd.read_csv(\"./train-data/sql_train.tsv\", sep='\\t')\n", "print(\"Total dataset examples: \" + str(len(df)))\n", "print(\"\\n\")\n", "\n", "# Test sampling\n", "sample = df.sample(n=1)\n", "print(sample[\"natural_query\"].values[0])\n", "print(sample[\"sql_query\"].values[0])\n", "print(sample[\"result\"].values[0])" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Load pre-trained DeepSeek model using transformers and pytorch packages" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "c:\\Users\\Dean\\AppData\\Local\\Programs\\Python\\Python312\\Lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", " from .autonotebook import tqdm as notebook_tqdm\n" ] } ], "source": [ "from transformers import AutoTokenizer, AutoModelForCausalLM\n", "import torch\n", "\n", "# Set device to cuda if available, otherwise CPU\n", "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", "\n", "# Load model and tokenizer\n", "tokenizer = AutoTokenizer.from_pretrained(\"./deepseek-coder-1.3b-instruct\")\n", "model = AutoModelForCausalLM.from_pretrained(\"./deepseek-coder-1.3b-instruct\", torch_dtype=torch.bfloat16, device_map=device) " ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Create prompt to setup the model for better performance" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [], "source": [ "input_text = \"\"\"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 game statistics, linked to the game table via game_id.\n", "CREATE TABLE IF NOT EXISTS \"other_stats\" (\n", " \"game_id\" TEXT, -- Unique game identifier (links to \"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 in the game\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 in the game\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", " \"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 in the game\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", "\n", "\n", "Query Guidelines\n", "Use team_name_home and team_name_away to match teams.\n", "\n", "To filter by season, use season_id = '2YYYY'.\n", "\n", "Example: To get games from 2005, use season_id = '22005'. To get games from 1972, use season_id = \"21972\". To get games from 2015, use season_id = \"22015\".\n", "\n", "The game_id column links the game and other_stats tables.\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) \n", "FROM game \n", "WHERE team_name_home = 'Los Angeles Lakers';\n", "\n", "Request:\n", "\"How many points did the Miami Heat score on January 10, 2010?\"\n", "SQLite:\n", "SELECT team_name_home, pts_home, team_name_away, pts_away \n", "FROM game \n", "WHERE DATE(game_date) = '2010-01-10' \n", "AND (team_name_home = 'Miami Heat' OR team_name_away = 'Miami Heat');\n", "\n", "Request:\n", "\"Which team won the most home games in the 2000 season?\"\n", "SQLite:\n", "SELECT team_name_home, COUNT(*) AS wins\n", "FROM game\n", "WHERE wl_home = 'W' AND season_id = '22000'\n", "GROUP BY team_name_home\n", "ORDER BY wins DESC\n", "LIMIT 1;\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 question: \"\"\"" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Test model performance on a single example" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "c:\\Users\\Dean\\AppData\\Local\\Programs\\Python\\Python312\\Lib\\site-packages\\transformers\\generation\\configuration_utils.py:634: UserWarning: `do_sample` is set to `False`. However, `top_p` is set to `0.95` -- this flag is only used in sample-based generation modes. You should set `do_sample=True` or unset `top_p`.\n", " warnings.warn(\n", "The attention mask and the pad token id were not set. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.\n", "Setting `pad_token_id` to `eos_token_id`:32021 for open-end generation.\n", "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\\transformers\\integrations\\sdpa_attention.py:53: UserWarning: 1Torch was not compiled with flash attention. (Triggered internally at C:\\actions-runner\\_work\\pytorch\\pytorch\\builder\\windows\\pytorch\\aten\\src\\ATen\\native\\transformers\\cuda\\sdp_utils.cpp:555.)\n", " attn_output = torch.nn.functional.scaled_dot_product_attention(\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "SQLite:\n", "SELECT MAX(largest_lead_home) \n", "FROM other_stats \n", "WHERE team_name_home = 'Golden State Warriors' AND season_id = '22018';\n", "\n" ] } ], "source": [ "# Create message with sample query and run model\n", "message=[{ 'role': 'user', 'content': input_text + sample[\"natural_query\"].values[0]}]\n", "inputs = tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors=\"pt\").to(model.device)\n", "outputs = model.generate(inputs, max_new_tokens=512, do_sample=False, top_k=50, top_p=0.95, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id)\n", "\n", "# Print output\n", "query_output = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)\n", "print(query_output)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# Test sample output on sqlite3 database" ] }, { "cell_type": "code", "execution_count": 5, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "cleaned\n" ] }, { "ename": "OperationalError", "evalue": "no such column: team_name_home", "output_type": "error", "traceback": [ "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m", "\u001b[1;31mOperationalError\u001b[0m Traceback (most recent call last)", "Cell \u001b[1;32mIn[5], line 15\u001b[0m\n\u001b[0;32m 13\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[0;32m 14\u001b[0m query \u001b[38;5;241m=\u001b[39m query_output\n\u001b[1;32m---> 15\u001b[0m \u001b[43mcursor\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mexecute\u001b[49m\u001b[43m(\u001b[49m\u001b[43mquery\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m 16\u001b[0m rows \u001b[38;5;241m=\u001b[39m cursor\u001b[38;5;241m.\u001b[39mfetchall()\n\u001b[0;32m 17\u001b[0m \u001b[38;5;28;01mfor\u001b[39;00m row \u001b[38;5;129;01min\u001b[39;00m rows:\n", "\u001b[1;31mOperationalError\u001b[0m: no such column: team_name_home" ] } ], "source": [ "import sqlite3 as sql\n", "\n", "# Create connection to sqlite3 database\n", "connection = sql.connect('./nba-data/nba.sqlite')\n", "cursor = connection.cursor()\n", "\n", "# Execute query from model output and print result\n", "if query_output[0:7] == \"SQLite:\":\n", " print(\"cleaned\")\n", " query = query_output[7:]\n", "elif query_output[0:4] == \"SQL:\":\n", " query = query_output[4:]\n", "else:\n", " query = query_output\n", "cursor.execute(query)\n", "rows = cursor.fetchall()\n", "for row in rows:\n", " print(row)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Create function to compare output to ground truth result from examples" ] }, { "cell_type": "code", "execution_count": 17, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "The attention mask and the pad token id were not set. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.\n", "Setting `pad_token_id` to `eos_token_id`:32021 for open-end generation.\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "How many games had at least one team with 30+ assists?\n", "SELECT COUNT(*) FROM game WHERE ast_home >= 30 OR ast_away >= 30;\n", "11305\n", "SQLite:\n", "SELECT COUNT(*) \n", "FROM game \n", "WHERE ast_home >= 30 OR ast_away >= 30;\n", "\n", "[(11305,)]\n", "SQL matched? True\n", "Result matched? True\n" ] } ], "source": [ "def compare_result(sample_query, sample_result, query_output):\n", " # Clean model output to only have the query output\n", " if query_output[0:7] == \"SQLite:\":\n", " query = query_output[7:]\n", " elif query_output[0:4] == \"SQL:\":\n", " query = query_output[4:]\n", " else:\n", " query = query_output\n", " \n", " # Try to execute query, if it fails, then this is a failure of the model\n", " try:\n", " # Execute query and obtain result\n", " cursor.execute(query)\n", " rows = cursor.fetchall()\n", "\n", " # Strip all whitespace before comparing queries since there may be differences in spacing, newlines, tabs, etc.\n", " query = query.replace(\" \", \"\").replace(\"\\n\", \"\").replace(\"\\t\", \"\")\n", " sample_query = sample_query.replace(\" \", \"\").replace(\"\\n\", \"\").replace(\"\\t\", \"\")\n", " query_match = (query == sample_query)\n", "\n", " # Check if this is a multi-line query\n", " if \"|\" in sample_result or \"(\" in sample_result:\n", " if \"(\" in sample_result:\n", " sample_result = sample_result.replace(\"(\", \"\").replace(\")\", \"\")\n", " result_list = sample_result.split(\",\") \n", " else:\n", " result_list = sample_result.split(\"|\") \n", "\n", " for i in range(len(result_list)):\n", " result_list[i] = str(result_list[i]).strip()\n", " result = False\n", " for row in rows:\n", " for r in row:\n", " if str(r) in result_list:\n", " return query_match, True\n", " if len(rows) == 1:\n", " for r in rows[0]:\n", " if r == str(len(result_list)):\n", " return query_match, True\n", " return query_match, result\n", " else:\n", " print(rows)\n", " result = False\n", " for row in rows:\n", " for r in row:\n", " if str(r) in str(sample_result):\n", " return query_match, True\n", "\n", " # Compare results and return\n", " return query_match, result\n", " except:\n", " return False, False\n", "\n", "# Obtain sample\n", "sample = df.sample(n=1)\n", "print(sample[\"natural_query\"].values[0])\n", "print(sample[\"sql_query\"].values[0])\n", "print(sample[\"result\"].values[0])\n", "\n", "# Create message with sample query and run model\n", "message=[{ 'role': 'user', 'content': input_text + sample[\"natural_query\"].values[0]}]\n", "inputs = tokenizer.apply_chat_template(message, add_generation_prompt=True, return_tensors=\"pt\").to(model.device)\n", "outputs = model.generate(inputs, max_new_tokens=512, do_sample=False, top_k=50, top_p=0.95, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id)\n", "\n", "# Print output\n", "query_output = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)\n", "print(query_output)\n", "\n", "result = compare_result(sample[\"sql_query\"].values[0], sample[\"result\"].values[0], query_output)\n", "print(\"SQL matched? \" + str(result[0]))\n", "print(\"Result matched? \" + str(result[1]))" ] } ], "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 }