diff --git a/docs/blog/posts/distilation-part1.md b/docs/blog/posts/distilation-part1.md new file mode 100644 index 0000000..22508e7 --- /dev/null +++ b/docs/blog/posts/distilation-part1.md @@ -0,0 +1,84 @@ +--- +draft: False +date: 2023-10-17 +tags: + - RAG + - Finetuning +--- + +# Introduction to `Instructions` from `Instructor`, finetuning from Python functions. + +The core philosophy with the `instructor` library is to make language models backwards compatible with existing code. By adding Pydantic in the mix we're able to easily work with LLMs without much worry. + +However, many times, a single function isn't just one LLM call. After the results are returned theres [validation](/docs/validation.md), some additional processing and formatting before you `return` the result. + +But the promise of LLMs is that they can do all of this in one go. So how do we get there? Finetuning end to end is a great tool for enhancing language models. Instructor uses type hints via Pydantic to maintain backward compatibility. Distillation focuses on fine-tuning language models to imitate specific functions. + +## Challenges in Fine-tuning + +Fine-tuning a model isn't as straightforward as just writing `def f(a, b): return a * b` to teach a model three-digit multiplication. Substantial data preparation is required, making logging for data collection cumbersome. Luckily OpenAI not only provides a fine-tuning script but also one for function calling which simplies the process backed to structured outputs! More over, the finetune allows us to avoid passing the schema to the model, resulting in less tokens being used! + +## Role of Instructor in Easing the Process + +The feature `from instructor import Instructions` simplifies this. It decorates Python functions that return Pydantic objects, automatically creating a fine-tuning dataset when provided a handler for logging. This allows you to finetune a model to imitate a function's behavior. + +## How to Use Instructor's Distillation Feature + +Here's an example to illustrate its use: + +```python +import logging +from pydantic import BaseModel +from instructor import Instructions + +logging.basicConfig(level=logging.INFO) + +instructions = Instructions( + name="three_digit_multiply", + finetune_format="messages", + log_handlers=[logging.FileHandler("math_finetunes.jsonl")] +) + +class Response(BaseModel): + a: int + b: int + result: int + +@instructions.distil +def fn(a: int, b: int) -> Response: + resp = a + b + return Response(a=a, b=b, result=resp) + +for _ in range(10): + a = random.randint(100, 999) + b = random.randint(100, 999) + print(fn(a, b)) +``` + +## Why Instructor and Distillation are Useful + +1. Streamlines complex functions with validations, making them more efficient. +2. Facilitates the integration of classical machine learning with language models. + +By understanding and leveraging these capabilities, you can create powerful, fine-tuned language models with ease. To learn more about how to use the file to finetune a model, check out the [cli](/docs/cli/finetune.md) + +## Next Steps + +This post is mostly a peek of what I've been working on this week. Once we have a model trained I'd like to be able to dynamically swap the implemetnation of a function with a model. This would allow us to do things like: + +```python +from instructor import Instructions + +instructions = Instructions( + name="three_digit_multiply", +) + +@instructions.distil(model='gpt-3.5-turbo:finetuned', swap=True) +def fn(a: int, b: int) -> Response: + resp = a + b + return Response(a=a, b=b, result=resp) +``` + +Now we can swap out the implementation of `fn` with calling the finetuned model, since we know the response type is still `Response` we can use instructor behind the scenes and have it be backwards compatible with the existing code. + +This is a powerful idea, and I'm excited to see where it goes. \ No newline at end of file diff --git a/docs/distilation.md b/docs/distilation.md new file mode 100644 index 0000000..796f6e6 --- /dev/null +++ b/docs/distilation.md @@ -0,0 +1,93 @@ +# Distilling python functions into LLM + +`Instructions` from the `Instructor` library offers a seamless way to make language models backward compatible with existing Python functions. By employing Pydantic type hints, it not only ensures compatibility but also facilitates fine-tuning language models to emulate these functions end-to-end. + +## The Challenges in Function-Level Fine-Tuning + +Unlike simple script-level fine-tuning, replicating the behavior of a Python function in a language model involves intricate data preparation. For instance, teaching a model to execute three-digit multiplication is not as trivial as implementing `def f(a, b): return a * b`. OpenAI's fine-tuning script coupled with their function calling utility provides a structured output, thereby simplifying the data collection process. Additionally, this eliminates the need for passing the schema to the model, thus conserving tokens. + +## The Role of `Instructions` in Simplifying the Fine-Tuning Process + +By using `Instructions`, you can annotate a Python function that returns a Pydantic object, thereby automating the dataset creation for fine-tuning. A handler for logging is all that's needed to build this dataset. + +## How to Implement `Instructions` in Your Code + +Here's a step-by-step example: + +```python +import logging +from pydantic import BaseModel +from instructor import Instructions + +logging.basicConfig(level=logging.INFO) + +instructions = Instructions( + name="three_digit_multiply", + finetune_format="messages", + log_handlers=[logging.FileHandler("math_finetunes.jsonl")] +) + +class Response(BaseModel): + a: int + b: int + result: int + +@instructions.distil +def fn(a: int, b: int) -> Response: + resp = a + b + return Response(a=a, b=b, result=resp) +``` + +## Custom Log Handlers for Data Collection + +While the example above uses a file-based log handler, you can easily extend this to custom log handlers for different storage solutions. The following skeleton code illustrates how to create a log handler for an S3 bucket: + +```python +import logging +import boto3 + +class S3LogHandler(logging.Handler): + def __init__(self, bucket, key): + logging.Handler.__init__(self) + self.bucket = bucket + self.key = key + + def emit(self, record): + s3 = boto3.client('s3') + log_entry = self.format(record) + s3.put_object(Body=log_entry, Bucket=self.bucket, Key=self.key) +``` + +You can add this custom log handler to `Instructions` as shown: + +```python +instructions = Instructions( + name="three_digit_multiply", + finetune_format="messages", + log_handlers=[S3LogHandler(bucket='your-bucket', key='your-key')] +) +``` + +## Why `Instructions` is a Game-Changer + +1. It condenses complex, multi-step functions with validations into a single fine-tuned model. +2. It integrates language models with classical machine learning seamlessly. + +## Next Steps and Future Scope + +Going forward, the aim is to dynamically switch between the Python function and its fine-tuned model representation. This could look like: + +```python +from instructor import Instructions + +instructions = Instructions( + name="three_digit_multiply", +) + +@instructions.distil(model='gpt-3.5-turbo:finetuned', swap=True) +def fn(a: int, b: int) -> Response: + resp = a + b + return Response(a=a, b=b, result=resp) +``` + +This dynamic switching retains backward compatibility while improving efficiency, opening up exciting avenues for future developments. \ No newline at end of file diff --git a/mkdocs.yml b/mkdocs.yml index d386c21..447f7a6 100644 --- a/mkdocs.yml +++ b/mkdocs.yml @@ -56,6 +56,7 @@ nav: - Introduction: - Getting Started: 'index.md' - Prompt Engineering Tips: 'tips/index.md' + - Distillation: 'distillation.md' - Helpers: - Reasking and Validation Overview: "reask_validation.md" - Multiple Extractions: "multitask.md"