How to fine-tune chat models

OpenAI Logo
Simón Fishman
Aug 22, 2023
Open in Github

This notebook provides a step-by-step guide for our new gpt-3.5-turbo fine-tuning. We'll perform entity extraction using the RecipeNLG dataset, which provides various recipes and a list of extracted generic ingredients for each. This is a common dataset for named entity recognition (NER) tasks.

We will go through the following steps:

  1. Setup: Loading our dataset and filtering down to one domain to fine-tune on.
  2. Data preparation: Preparing your data for fine-tuning by creating training and validation examples, and uploading them to the Files endpoint.
  3. Fine-tuning: Creating your fine-tuned model.
  4. Inference: Using your fine-tuned model for inference on new inputs.

By the end of this you should be able to train, evaluate and deploy a fine-tuned gpt-3.5-turbo model.

For more information on fine-tuning, you can refer to our documentation guide, API reference or blog post

# make sure to use the latest version of the openai python package
!pip install --upgrade openai 
import json
import openai
import os
import pandas as pd
from pprint import pprint

OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "")

Fine-tuning works best when focused on a particular domain. It's important to make sure your dataset is both focused enough for the model to learn, but general enough that unseen examples won't be missed. Having this in mind, we have extracted a subset from the RecipesNLG dataset to only contain documents from www.cookbooks.com.

# Read in the dataset we'll use for this task.
# This will be the RecipesNLG dataset, which we've cleaned to only contain documents from www.cookbooks.com
recipe_df = pd.read_csv("data/cookbook_recipes_nlg_10k.csv")

recipe_df.head()
title ingredients directions link source NER
0 No-Bake Nut Cookies ["1 c. firmly packed brown sugar", "1/2 c. eva... ["In a heavy 2-quart saucepan, mix brown sugar... www.cookbooks.com/Recipe-Details.aspx?id=44874 www.cookbooks.com ["brown sugar", "milk", "vanilla", "nuts", "bu...
1 Jewell Ball'S Chicken ["1 small jar chipped beef, cut up", "4 boned ... ["Place chipped beef on bottom of baking dish.... www.cookbooks.com/Recipe-Details.aspx?id=699419 www.cookbooks.com ["beef", "chicken breasts", "cream of mushroom...
2 Creamy Corn ["2 (16 oz.) pkg. frozen corn", "1 (8 oz.) pkg... ["In a slow cooker, combine all ingredients. C... www.cookbooks.com/Recipe-Details.aspx?id=10570 www.cookbooks.com ["frozen corn", "cream cheese", "butter", "gar...
3 Chicken Funny ["1 large whole chicken", "2 (10 1/2 oz.) cans... ["Boil and debone chicken.", "Put bite size pi... www.cookbooks.com/Recipe-Details.aspx?id=897570 www.cookbooks.com ["chicken", "chicken gravy", "cream of mushroo...
4 Reeses Cups(Candy) ["1 c. peanut butter", "3/4 c. graham cracker ... ["Combine first four ingredients and press in ... www.cookbooks.com/Recipe-Details.aspx?id=659239 www.cookbooks.com ["peanut butter", "graham cracker crumbs", "bu...

Data preparation

We'll begin by preparing our data. When fine-tuning with the ChatCompletion format, each training example is a simple list of messages. For example, an entry could look like:

[{'role': 'system',
  'content': 'You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided.'},

 {'role': 'user',
  'content': 'Title: No-Bake Nut Cookies\n\nIngredients: ["1 c. firmly packed brown sugar", "1/2 c. evaporated milk", "1/2 tsp. vanilla", "1/2 c. broken nuts (pecans)", "2 Tbsp. butter or margarine", "3 1/2 c. bite size shredded rice biscuits"]\n\nGeneric ingredients: '},

 {'role': 'assistant',
  'content': '["brown sugar", "milk", "vanilla", "nuts", "butter", "bite size shredded rice biscuits"]'}]

During the training process this conversation will be split, with the final entry being the completion that the model will produce, and the remainder of the messages acting as the prompt. Consider this when building your training examples - if your model will act on multi-turn conversations, then please provide representative examples so it doesn't perform poorly when the conversation starts to expand.

Please note that currently there is a 4096 token limit for each training example. Anything longer than this will be truncated at 4096 tokens.

training_data = []

system_message = "You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided."

def create_user_message(row):
    return f"""Title: {row['title']}\n\nIngredients: {row['ingredients']}\n\nGeneric ingredients: """

def prepare_example_conversation(row):
    messages = []
    messages.append({"role": "system", "content": system_message})

    user_message = create_user_message(row)
    messages.append({"role": "user", "content": user_message})

    messages.append({"role": "assistant", "content": row["NER"]})

    return {"messages": messages}

pprint(prepare_example_conversation(recipe_df.iloc[0]))
{'messages': [{'content': 'You are a helpful recipe assistant. You are to '
                          'extract the generic ingredients from each of the '
                          'recipes provided.',
               'role': 'system'},
              {'content': 'Title: No-Bake Nut Cookies\n'
                          '\n'
                          'Ingredients: ["1 c. firmly packed brown sugar", '
                          '"1/2 c. evaporated milk", "1/2 tsp. vanilla", "1/2 '
                          'c. broken nuts (pecans)", "2 Tbsp. butter or '
                          'margarine", "3 1/2 c. bite size shredded rice '
                          'biscuits"]\n'
                          '\n'
                          'Generic ingredients: ',
               'role': 'user'},
              {'content': '["brown sugar", "milk", "vanilla", "nuts", '
                          '"butter", "bite size shredded rice biscuits"]',
               'role': 'assistant'}]}

Let's now do this for a subset of the dataset to use as our training data. You can begin with even 30-50 well-pruned examples. You should see performance continue to scale linearly as you increase the size of the training set, but your jobs will also take longer.

# use the first 100 rows of the dataset for training
training_df = recipe_df.loc[0:100]

# apply the prepare_example_conversation function to each row of the training_df
training_data = training_df.apply(prepare_example_conversation, axis=1).tolist()

for example in training_data[:5]:
    print(example)
{'messages': [{'role': 'system', 'content': 'You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided.'}, {'role': 'user', 'content': 'Title: No-Bake Nut Cookies\n\nIngredients: ["1 c. firmly packed brown sugar", "1/2 c. evaporated milk", "1/2 tsp. vanilla", "1/2 c. broken nuts (pecans)", "2 Tbsp. butter or margarine", "3 1/2 c. bite size shredded rice biscuits"]\n\nGeneric ingredients: '}, {'role': 'assistant', 'content': '["brown sugar", "milk", "vanilla", "nuts", "butter", "bite size shredded rice biscuits"]'}]}
{'messages': [{'role': 'system', 'content': 'You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided.'}, {'role': 'user', 'content': 'Title: Jewell Ball\'S Chicken\n\nIngredients: ["1 small jar chipped beef, cut up", "4 boned chicken breasts", "1 can cream of mushroom soup", "1 carton sour cream"]\n\nGeneric ingredients: '}, {'role': 'assistant', 'content': '["beef", "chicken breasts", "cream of mushroom soup", "sour cream"]'}]}
{'messages': [{'role': 'system', 'content': 'You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided.'}, {'role': 'user', 'content': 'Title: Creamy Corn\n\nIngredients: ["2 (16 oz.) pkg. frozen corn", "1 (8 oz.) pkg. cream cheese, cubed", "1/3 c. butter, cubed", "1/2 tsp. garlic powder", "1/2 tsp. salt", "1/4 tsp. pepper"]\n\nGeneric ingredients: '}, {'role': 'assistant', 'content': '["frozen corn", "cream cheese", "butter", "garlic powder", "salt", "pepper"]'}]}
{'messages': [{'role': 'system', 'content': 'You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided.'}, {'role': 'user', 'content': 'Title: Chicken Funny\n\nIngredients: ["1 large whole chicken", "2 (10 1/2 oz.) cans chicken gravy", "1 (10 1/2 oz.) can cream of mushroom soup", "1 (6 oz.) box Stove Top stuffing", "4 oz. shredded cheese"]\n\nGeneric ingredients: '}, {'role': 'assistant', 'content': '["chicken", "chicken gravy", "cream of mushroom soup", "shredded cheese"]'}]}
{'messages': [{'role': 'system', 'content': 'You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided.'}, {'role': 'user', 'content': 'Title: Reeses Cups(Candy)  \n\nIngredients: ["1 c. peanut butter", "3/4 c. graham cracker crumbs", "1 c. melted butter", "1 lb. (3 1/2 c.) powdered sugar", "1 large pkg. chocolate chips"]\n\nGeneric ingredients: '}, {'role': 'assistant', 'content': '["peanut butter", "graham cracker crumbs", "butter", "powdered sugar", "chocolate chips"]'}]}

In addition to training data, we can also optionally provide validation data, which will be used to make sure that the model does not overfit your training set.

validation_df = recipe_df.loc[101:200]
validation_data = validation_df.apply(prepare_example_conversation, axis=1).tolist()

We then need to save our data as .jsonl files, with each line being one training example conversation.

def write_jsonl(data_list: list, filename: str) -> None:
    with open(filename, "w") as out:
        for ddict in data_list:
            jout = json.dumps(ddict) + "\n"
            out.write(jout)
training_file_name = "tmp_recipe_finetune_training.jsonl"
write_jsonl(training_data, training_file_name)

validation_file_name = "tmp_recipe_finetune_validation.jsonl"
write_jsonl(validation_data, validation_file_name)

This is what the first 5 lines of our training .jsonl file look like:

# print the first 5 lines of the training file
!head -n 5 tmp_recipe_finetune_training.jsonl
{"messages": [{"role": "system", "content": "You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided."}, {"role": "user", "content": "Title: No-Bake Nut Cookies\n\nIngredients: [\"1 c. firmly packed brown sugar\", \"1/2 c. evaporated milk\", \"1/2 tsp. vanilla\", \"1/2 c. broken nuts (pecans)\", \"2 Tbsp. butter or margarine\", \"3 1/2 c. bite size shredded rice biscuits\"]\n\nGeneric ingredients: "}, {"role": "assistant", "content": "[\"brown sugar\", \"milk\", \"vanilla\", \"nuts\", \"butter\", \"bite size shredded rice biscuits\"]"}]}
{"messages": [{"role": "system", "content": "You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided."}, {"role": "user", "content": "Title: Jewell Ball'S Chicken\n\nIngredients: [\"1 small jar chipped beef, cut up\", \"4 boned chicken breasts\", \"1 can cream of mushroom soup\", \"1 carton sour cream\"]\n\nGeneric ingredients: "}, {"role": "assistant", "content": "[\"beef\", \"chicken breasts\", \"cream of mushroom soup\", \"sour cream\"]"}]}
{"messages": [{"role": "system", "content": "You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided."}, {"role": "user", "content": "Title: Creamy Corn\n\nIngredients: [\"2 (16 oz.) pkg. frozen corn\", \"1 (8 oz.) pkg. cream cheese, cubed\", \"1/3 c. butter, cubed\", \"1/2 tsp. garlic powder\", \"1/2 tsp. salt\", \"1/4 tsp. pepper\"]\n\nGeneric ingredients: "}, {"role": "assistant", "content": "[\"frozen corn\", \"cream cheese\", \"butter\", \"garlic powder\", \"salt\", \"pepper\"]"}]}
{"messages": [{"role": "system", "content": "You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided."}, {"role": "user", "content": "Title: Chicken Funny\n\nIngredients: [\"1 large whole chicken\", \"2 (10 1/2 oz.) cans chicken gravy\", \"1 (10 1/2 oz.) can cream of mushroom soup\", \"1 (6 oz.) box Stove Top stuffing\", \"4 oz. shredded cheese\"]\n\nGeneric ingredients: "}, {"role": "assistant", "content": "[\"chicken\", \"chicken gravy\", \"cream of mushroom soup\", \"shredded cheese\"]"}]}
{"messages": [{"role": "system", "content": "You are a helpful recipe assistant. You are to extract the generic ingredients from each of the recipes provided."}, {"role": "user", "content": "Title: Reeses Cups(Candy)  \n\nIngredients: [\"1 c. peanut butter\", \"3/4 c. graham cracker crumbs\", \"1 c. melted butter\", \"1 lb. (3 1/2 c.) powdered sugar\", \"1 large pkg. chocolate chips\"]\n\nGeneric ingredients: "}, {"role": "assistant", "content": "[\"peanut butter\", \"graham cracker crumbs\", \"butter\", \"powdered sugar\", \"chocolate chips\"]"}]}

Upload files

You can now upload the files to our Files endpoint to be used by the fine-tuned model.

training_response = openai.File.create(
    file=open(training_file_name, "rb"), purpose="fine-tune"
)
training_file_id = training_response["id"]

validation_response = openai.File.create(
    file=open(validation_file_name, "rb"), purpose="fine-tune"
)
validation_file_id = validation_response["id"]

print("Training file ID:", training_file_id)
print("Validation file ID:", validation_file_id)
Training file ID: file-jcdvNl27iuBMZfwi4q30IIka
Validation file ID: file-O144OIHkZ1xjB32ednBmbOXP

Fine-tuning

Now we can create our fine-tuning job with the generated files and an optional suffix to identify the model. The response will contain an id which you can use to retrieve updates on the job.

Note: The files have to first be processed by our system, so you might get a File not ready error. In that case, simply retry a few minutes later.

response = openai.FineTuningJob.create(
    training_file=training_file_id,
    validation_file=validation_file_id,
    model="gpt-3.5-turbo",
    suffix="recipe-ner",
)

job_id = response["id"]

print("Job ID:", response["id"])
print("Status:", response["status"])
Job ID: ftjob-9xVzrp0Oem9rWj2UFWDcFLqT
Status: created

Check job status

You can make a GET request to the https://api.openai.com/v1/alpha/fine-tunes endpoint to list your alpha fine-tune jobs. In this instance you'll want to check that the ID you got from the previous step ends up as status: succeeded.

Once it is completed, you can use the result_files to sample the results from the validation set (if you uploaded one), and use the ID from the fine_tuned_model parameter to invoke your trained model.

response = openai.FineTuningJob.retrieve(job_id)

print("Job ID:", response["id"])
print("Status:", response["status"])
print("Trained Tokens:", response["trained_tokens"])

Job ID: ftjob-9xVzrp0Oem9rWj2UFWDcFLqT
Status: running
Trained Tokens: None

We can track the progress of the fine-tune with the events endpoint. You can rerun the cell below a few times until the fine-tune is ready.

response = openai.FineTuningJob.list_events(id=job_id, limit=50)

events = response["data"]
events.reverse()

for event in events:
    print(event["message"])
Created fine-tune: ftjob-9xVzrp0Oem9rWj2UFWDcFLqT
Fine tuning job started
Step 10: training loss=2.41
Step 20: training loss=0.06
Step 30: training loss=0.38
Step 40: training loss=0.09
Step 50: training loss=0.19
Step 60: training loss=0.10
Step 70: training loss=0.00
Step 80: training loss=0.01
Step 90: training loss=0.72
Step 100: training loss=0.13
Step 110: training loss=0.15
Step 120: training loss=0.00
Step 130: training loss=0.47
Step 140: training loss=0.00
Step 150: training loss=0.10
Step 160: training loss=0.06
Step 170: training loss=0.03
Step 180: training loss=0.08
Step 190: training loss=0.04
Step 200: training loss=0.01
Step 210: training loss=0.03
Step 220: training loss=0.00
Step 230: training loss=0.08
Step 240: training loss=0.00
Step 250: training loss=0.00
Step 260: training loss=0.00
Step 270: training loss=0.00
Step 280: training loss=0.00
Step 290: training loss=0.00
Step 300: training loss=0.60
New fine-tuned model created: ft:gpt-3.5-turbo-0613:openai:recipe-ner:7qTvyJ81
Fine-tuning job successfully completed

Now that it's done, we can get a fine-tuned model ID from the job:

response = openai.FineTuningJob.retrieve(job_id)
fine_tuned_model_id = response["fine_tuned_model"]

if fine_tuned_model_id is None: 
    raise RuntimeError("Fine-tuned model ID not found. Your job has likely not been completed yet.")

print("Fine-tuned model ID:", fine_tuned_model_id)
Fine-tuned model ID: ft:gpt-3.5-turbo-0613:openai:recipe-ner:7qTvyJ81

The last step is to use your fine-tuned model for inference. Similar to the classic FineTuning, you simply call ChatCompletions with your new fine-tuned model name filling the model parameter.

test_df = recipe_df.loc[201:300]
test_row = test_df.iloc[0]
test_messages = []
test_messages.append({"role": "system", "content": system_message})
user_message = create_user_message(test_row)
test_messages.append({"role": "user", "content": create_user_message(test_row)})

pprint(test_messages)
[{'content': 'You are a helpful recipe assistant. You are to extract the '
             'generic ingredients from each of the recipes provided.',
  'role': 'system'},
 {'content': 'Title: Pancakes\n'
             '\n'
             'Ingredients: ["1 c. flour", "1 tsp. soda", "1 tsp. salt", "1 '
             'Tbsp. sugar", "1 egg", "3 Tbsp. margarine, melted", "1 c. '
             'buttermilk"]\n'
             '\n'
             'Generic ingredients: ',
  'role': 'user'}]

response = openai.ChatCompletion.create(
    model=fine_tuned_model_id, messages=test_messages, temperature=0, max_tokens=500
)
print(response["choices"][0]["message"]["content"])
["flour", "soda", "salt", "sugar", "egg", "margarine", "buttermilk"]

Conclusion

Congratulations, you are now ready to fine-tune your own models using the ChatCompletion format! We look forward to seeing what you build