epec254 / dspy_examples

Overview

This is Eric's playground with DSPy.

It includes sample DSPy modules for GSM8K and MMLU. These are fully self contained e.g., you can run them by going python ./name_of_file.py and it just works.

Tested on Python 3.11. Install pip install mlflow dpsy-ai.

GSM8k

1. Open dpsy_gsm8k.py
2. Add your OpenAI key to line 18 os.environ["OPENAI_API_KEY"] = ...
3. Edit lines 169-171 if you want to run on a subset of the dataset first to test
4. Start the local MLflow server with mlflow server --host 127.0.0.1 --port 8080
5. Execute python ./dpsy_gsm8k.py
6. View the runs in MLflow and/or via the console
7. To use the logged MLflow model, see load_gsm8k_from_mlflow.py.

Optionally, adjust ENABLE_ARIZE_TRACING = true if you want to see trace logs in Phoenix. You'll need to:

• Install the packages pip install arize-phoenix openinference-instrumentation-dspy opentelemetry-exporter-otlp
• Start their container docker run -p 6006:6006 arizephoenix/phoenix:latest

MMLU

1. Open dpsy_gsm8k.py
2. Add your OpenAI key to line 18 os.environ["OPENAI_API_KEY"] = ...
3. Edit lines 193-195 if you want to run on a subset of the dataset first to test
4. Start the local MLflow server with mlflow server --host 127.0.0.1 --port 8080
5. Execute python ./dpsy_mmlu.py
6. View the runs in MLflow and/or via the console

If you prefer to re-ETL the MMLU data, you can run download_mmlu.py. However, the results of this download are saved in mmlu_*.json.

Azure OpenAI

Azure OpenAI's content safety filter blocks DSPy's generated queries so this isn't fully tested. azureopenai.py is a working client for Azure OpenAI since the default DSPy implementation doesn't work. TODO: Contribute this back and clean up the depedencies between the DSPy AzureOpenAI library and the client code.

from azureopenai import AzureOpenAI
os.environ["AZURE_OPENAI_API_KEY"] = "<your-key-here>"
os.environ["AZURE_OPENAI_ENDPOINT"] = "https://your-url.openai.azure.com/"

model = AzureOpenAI(
     api_provider="azure",
     model_type="chat",
     api_version="2023-07-01-preview", # version from Azure 2023-07-01-preview
     model="deployment_name",  # deployment_name from Azure - often the same as the model name e.g., chat-3.5 but not necessarly
)
dspy.configure(lm=model)

TODO

• Auto generate the PyFunc wrapper from the DSPy signatures
• Refactor out the common boiler plate code between GSM8k and MMLU