Notebooks
O
OpenAI Chat Finetuning Data Prep

Chat Finetuning Data Prep

Data preparation and analysis for chat model fine-tuning

This notebook serves as a tool to preprocess and analyze the chat dataset used for fine-tuning a chat model. It checks for format errors, provides basic statistics, and estimates token counts for fine-tuning costs. The method shown here corresponds to the current fine-tuning method for gpt-3.5-turbo. See legacy fine-tuning for models like babbage-002 and davinci-002.

[1]

Data loading

We first load the chat dataset from an example JSONL file.

[2]
Num examples: 5
First example:
{'role': 'system', 'content': 'You are a happy assistant that puts a positive spin on everything.'}
{'role': 'user', 'content': 'I fell off my bike today.'}
{'role': 'assistant', 'content': "It's great that you're getting exercise outdoors!"}

Format validation

We can perform a variety of error checks to validate that each conversation in the dataset adheres to the format expected by the fine-tuning API. Errors are categorized based on their nature for easier debugging.

  1. Data Type Check: Checks whether each entry in the dataset is a dictionary (dict). Error type: data_type.
  2. Presence of Message List: Checks if a messages list is present in each entry. Error type: missing_messages_list.
  3. Message Keys Check: Validates that each message in the messages list contains the keys role and content. Error type: message_missing_key.
  4. Unrecognized Keys in Messages: Logs if a message has keys other than role, content, weight, function_call, and name. Error type: message_unrecognized_key.
  5. Role Validation: Ensures the role is one of "system", "user", or "assistant". Error type: unrecognized_role.
  6. Content Validation: Verifies that content has textual data and is a string. Error type: missing_content.
  7. Assistant Message Presence: Checks that each conversation has at least one message from the assistant. Error type: example_missing_assistant_message.

The code below performs these checks, and outputs counts for each type of error found are printed. This is useful for debugging and ensuring the dataset is ready for the next steps.

[3]
No errors found

Token Counting Utilities

Lets define a few helpful utilities to be used in the rest of the notebook.

[4]

Data Warnings and Token Counts

With some lightweight analysis we can identify potential issues in the dataset, like missing messages, and provide statistical insights into message and token counts.

  1. Missing System/User Messages: Counts the number of conversations missing a "system" or "user" message. Such messages are critical for defining the assistant's behavior and initiating the conversation.
  2. Number of Messages Per Example: Summarizes the distribution of the number of messages in each conversation, providing insight into dialogue complexity.
  3. Total Tokens Per Example: Calculates and summarizes the distribution of the total number of tokens in each conversation. Important for understanding fine-tuning costs.
  4. Tokens in Assistant's Messages: Calculates the number of tokens in the assistant's messages per conversation and summarizes this distribution. Useful for understanding the assistant's verbosity.
  5. Token Limit Warnings: Checks if any examples exceed the maximum token limit (16,385 tokens), as such examples will be truncated during fine-tuning, potentially resulting in data loss.
[5]
Num examples missing system message: 1
Num examples missing user message: 1

#### Distribution of num_messages_per_example:
min / max: 2, 9
mean / median: 3.8, 3.0
p5 / p95: 2.0, 6.6000000000000005

#### Distribution of num_total_tokens_per_example:
min / max: 26, 8032
mean / median: 1648.4, 45.0
p5 / p95: 26.8, 4863.6

#### Distribution of num_assistant_tokens_per_example:
min / max: 4, 8000
mean / median: 1610.2, 10.0
p5 / p95: 6.0, 4811.200000000001

0 examples may be over the 16,385 token limit, they will be truncated during fine-tuning

Cost Estimation

In this final section, we estimate the total number of tokens that will be used for fine-tuning, which allows us to approximate the cost. It is worth noting that the duration of the fine-tuning jobs will also increase with the token count.

[6]
Dataset has ~4306 tokens that will be charged for during training
By default, you'll train for 20 epochs on this dataset
By default, you'll be charged for ~86120 tokens

See https://openai.com/pricing to estimate total costs.