SFT of Qwen2.5-Coder-7B
Dataset is a combination of BIRD and Spider; answers were created through knowledge distillation with CoT prompting from Qwen2.5-Coder-32B, and filtered for correctness.

We increased the BIRD dev set performance from 51.1% to 60%.
Our model beats much larger universal models like GPT-4o and Gemini 1.5 Pro, demonstrating the effectiveness of black-box knowledge distillation for constrained domains.
After training, this checkpoint even exceeds the performance of its teacher model, Qwen2.5-Coder-32B, which can be achieved by filtering the teacher’s outputs for Quality

Prompt:

Single Prompt used throughout the Dataset so best used as follows:

sql

CREATE TABLE geographic\n(\n    city   TEXT not null\n        primary key,\n    county TEXT null,\n    region TEXT null\n) \n /* \n 2 example rows: \n SELECT * FROM geographic LIMIT 2; \n    city              county   region \nalameda      alameda county bay area \n  alamo contra costa county bay area  \n */\n\n <br>
CREATE TABLE generalinfo\n(\n    id_restaurant INTEGER         not null\n        primary key,\n    label         TEXT  null,\n    food_type     TEXT  null,\n    city          TEXT  null,\n    review        REAL null,\n    foreign key (city) references geographic(city)\n            on update cascade on delete cascade\n) \n /* \n 2 example rows: \n SELECT * FROM generalinfo LIMIT 2; \n id_restaurant                label     food_type          city review \n            1       sparky's diner 24 hour diner san francisco    2.3 \n            2 kabul afghan cuisine       afghani    san carlos    3.8  \n */\n\nCREATE TABLE location\n(\n    id_restaurant INTEGER        not null\n        primary key,\n    street_num    INTEGER          null,\n    street_name   TEXT null,\n    city          TEXT null,\n    foreign key (city) references geographic (city)\n            on update cascade on delete cascade,\n    foreign key (id_restaurant) references generalinfo (id_restaurant)\n            on update cascade on delete cascade\n) \n /* \n 2 example rows: \n SELECT * FROM location LIMIT 2; \n id_restaurant street_num    street_name          city \n            1        242      church st san francisco \n            2        135 el camino real    san carlos  \n */\n\n  <br>
-- External Knowledge: Atlantic Ave refers to street_name = 'atlantic ave'; rating refers to review\n  <br>
-- Using valid SQLite and understanding External Knowledge, answer the following question for the tables provided above.\n  <br>
-- What is the rating of each restaurant reviews on Atlantic Ave?\n <br>
Generate the SQL after thinking step by step:\n  <br>

python

def bird_gpt_template_no_format(question, commonsense, schema):
    return f"""{schema}

-- External Knowledge: {commonsense}
-- Using valid SQLite and understanding External Knowledge, answer the following question for the tables provided above.
-- {question}
Generate the SQL after thinking step by step:
"""

Generate Schema:

python

def generate_schema_prompt(db_path, num_rows=None):
    # extract create ddls
    '''
    :param root_place:
    :param db_name:
    :return:
    '''
    full_schema_prompt_list = []
    conn = sqlite3.connect(db_path)
    # Create a cursor object
    cursor = conn.cursor()
    cursor.execute("SELECT name FROM sqlite_master WHERE type='table'")
    tables = cursor.fetchall()
    schemas = {}
    for table in tables:
        if table == 'sqlite_sequence':
            continue
        cursor.execute("SELECT sql FROM sqlite_master WHERE type='table' AND name='{}';".format(table[0]))
        create_prompt = cursor.fetchone()[0]
        schemas[table[0]] = create_prompt
        if num_rows:
            cur_table = table[0]
            if cur_table in ['order', 'by', 'group','transaction'] or ' ' in str(cur_table).strip() or '-' in str(cur_table).strip():
                cur_table = '"{}"'.format(cur_table)


            cursor.execute("SELECT * FROM {} LIMIT {}".format(cur_table, num_rows))
            column_names = [description[0] for description in cursor.description]
            values = cursor.fetchall()
            rows_prompt = nice_look_table(column_names=column_names, values=values)
            verbose_prompt = "/* \n {} example rows: \n SELECT * FROM {} LIMIT {}; \n {} \n */".format(num_rows,
                                                                                                       cur_table,
                                                                                                       num_rows,
                                                                                                       rows_prompt)
            schemas[table[0]] = "{} \n {}".format(create_prompt, verbose_prompt)

    for k, v in schemas.items():
        full_schema_prompt_list.append(v)

    schema_prompt = "\n\n".join(full_schema_prompt_list)

    return schema_prompt

System Prompt:

Default Qwen2.5 System Prompt

python

def preprocess_prompt(prompt):
    return f'''<|im_start|>system
You are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
'''

Generation Config:

No Sampling for best Performance

Restrictions:

only trained on SQLite Dialect
only trained in English
Did not care to keep any other Skills than Text to SQL

AlioLeuchtmann/ALIO-SQL-7B

This Model is deprecated we achieve +10% on the private test Set with a 2B Model, for more see here: https://huggingface.co/AlioLeuchtmann/qwen-3.5-sql-27B-Distill-2B

57% on BIRD private test Set (Qwen 2.5 Coder Models might have been trained on BIRD dev, which could explain the Performance Drop)

60% on BIRD dev greedy

70% on BIRD dev pass4

Model Details

Prompt:

Generate Schema:

System Prompt:

Generation Config:

Restrictions:

AlioLeuchtmann/
ALIO-SQL-7B