Add text generation task type

[bnativi]

Improvements

• Added a new evaluate_text_generation task type that calculates four new metrics. These include text comparison metrics, which compare a prediction string to a groundtruth string. These also include llm-guided metrics which only sometimes require a groundtruth. The metrics are:
  • AnswerRelevance (Q&A, llm-guided)
  • Bias (general text generation, llm-guided)
  • BLEU (text comparison)
  • Coherence (general text generation, llm-guided)
  • ContextRelevance (RAG, llm-guided)
  • Faithfulness (RAG, llm-guided)
  • Hallucination (RAG, llm-guided)
  • ROUGE (text comparison)
  • Toxicity (general text generation, llm-guided)
• Added text generation notebook with three example use cases (RAG, summarization and content generation).
• Added WrappedOpenAIClient and WrappedMistralAIClient to handle llm calls and llm-guided metric computations.
• Changed from alpine to slim.

Testing

• Added API functional and unit tests for the text generation metrics and llm clients.
• Added client side integration tests for text generation metrics.
• Added external integration tests to test the llm-guided metrics with OpenAI's API and Mistral's API. Because those API's do not give us fully deterministic control, the integration tests only check that valid metrics are returned and do not check the exact metric values.
  • These should only run on merge to main, and not on pushes to other branches.
    
  • They pass when run.
    
• If I purposely try to make them fail, say by setting the OPENAI_API_KEY to "", then they fail as expected.
  
  • The secret API keys should only be available to the external API integration tests and not to the rest of the integration tests. When I added a test to integration_tests/client/ that tries to make non-mocked OPENAI API calls, the test fails because the api key is not available for this test.
    

[ntlind]

the code, tests, and notebook all look good to me. I'm ready to approve once we figure out why the benchmarks are now failing.