Feature/implement the fuzz tests in robustness

[chakravarthik27]

Harness Setup:

from langtest import Harness 

harness = Harness(
    task="question-answering",
    model={
        "model": "llama3.1",
        "hub": "ollama",
        "type": "chat",
    },
    data={
        "data_source": "MedQA",
        "split": "test-tiny",
    },
    config={
        "tests": {
            "defaults": {
                "min_pass_rate": 0.5,
            },
            "clinical": {
                "medfuzz": {
                    "min_pass_rate": 0.1,
                    "attacker_llm": {
                        "model": "gpt-4o-mini",
                        "hub": "openai",
                        "type": "chat",
                    },                   
                }
            }
        }
    }
)

harness.generate().run().report()

[Copilot]

Pull Request Overview

This PR implements fuzz tests for robustness by introducing the MedFuzz feature. Key changes include:

• Refactoring CSV loading in utils.py to download remote files.
• Creating new LLM interaction classes (TargetLLM, AttackerLLM) and a MedFuzz class in clinical.py for processing clinical samples.
• Extending sample data types with HTML highlighting to display differences for MedFuzz samples.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
langtest/transform/utils.py	Refactored load_csv to download files; added TargetLLM and AttackerLLM classes.
langtest/transform/clinical.py	Introduced MedFuzz class and integrated LLM interactions for fuzz testing.
langtest/utils/custom_types/helpers.py	Added highlight_differences_both function for generating HTML diff highlights.
langtest/utils/custom_types/sample.py	Added MedFuzzSample subclass that overrides to_dict to incorporate HTML diff highlighting.

Comments suppressed due to low confidence (3)

langtest/transform/utils.py:574

• The new CSV downloading mechanism uses requests.get on the 'filepath' variable assuming it is a URL. Consider validating the input or adding error handling to ensure that non-URL file paths are managed appropriately.

# save the csv file into `~/.langtest/` directory

langtest/transform/utils.py:1052

• The error message here is unclear. Consider rephrasing it to clearly indicate the unsupported configuration in the LLM client.

raise TypeError("Unsupported hub and model and Only LLM")

langtest/transform/clinical.py:981

• Slicing the joined expected_results with [:1] might unintentionally truncate the value. Verify that this behavior is intended, or adjust to preserve the full expected result as needed.

med_sample.expected_results = "".join(map(str, med_sample.expected_results))[:1]