[GSK-1473] Enabling rewrite_prompt for giskard evaluator

[rabah-khalek]

• Refactor Giskard evaluator into utils and evaluation artifacts
• Re-introduce unwrapping of underlying object from pyfunc models
• test LLM scan detectors that didn't use to work

[linear]

GSK-1473 Enabling rewrite_prompt for giskard evaluator

see discussion in #1189

[rabah-khalek]

@andreybavt to make the review easier (since the change is blurred by the refactoring), here're the main modifications:

https://github.com/Giskard-AI/giskard/blob/31006aa2a84281fb7839dba03a2c958bcbe82def/python-client/giskard/integrations/mlflow/giskard_evaluator_utils.py#L30-L39

https://github.com/Giskard-AI/giskard/blob/31006aa2a84281fb7839dba03a2c958bcbe82def/python-client/giskard/integrations/mlflow/giskard_evaluator_utils.py#L79-L82

so now I unwrap the langchain model from the pyfunc model to be able to use rewrite_prompt and I tested it with:

models = ["text-ada-001"] 

import pandas as pd
import mlflow
import openai
import os
from langchain import PromptTemplate, LLMChain
from langchain.llms import OpenAI

df = pd.read_csv('https://raw.githubusercontent.com/sunnysai12345/News_Summary/master/news_summary_more.csv')

df_sample = pd.DataFrame(df["text"].sample(10, random_state=11))

os.environ["OPENAI_API_KEY"] = "XXX"
os.getenv("OPENAI_API_KEY")
openai.api_key = os.getenv("OPENAI_API_KEY")

prompt = PromptTemplate(template="Create a reader comment according to the following article summary: '{text}'",
                        input_variables=["text"])

for model in models:
    llm = OpenAI(openai_api_key=openai.api_key,
                 request_timeout=20,
                 max_retries=100,
                 temperature=0,
                 model_name=model)

    chain = LLMChain(prompt=prompt, llm=llm)

    evaluator_config = {"scan_config": {"params": {"text_perturbation": {"num_samples": 10},
                                                   "ethical_bias": {"num_samples": 10}}}}

    with mlflow.start_run(run_name = model):
        model_uri = mlflow.langchain.log_model(chain, "langchain").model_uri
        mlflow.evaluate(model=model_uri, model_type="text", data=df_sample, evaluators="giskard", evaluator_config=evaluator_config)

which outputs the following issues (that didn't use to work before):

Running detector LLMToxicityDetector…2023-07-27 19:05:51,229 pid:83686 MainThread giskard.scanner.logger ERROR    Detector LLMToxicityDetector failed with error: keywords must be strings
 0 issues detected. (Took 0:00:02.373041)
Running detector HarmfulnessDetector… 1 issues detected. (Took 0:00:13.804845)
Running detector GenderStereotypeDetector… 0 issues detected. (Took 0:00:22.148888)
Running detector MinorityStereotypeDetector… 1 issues detected. (Took 0:00:46.699300)

[mattbit]

@rabah-khalek why is the LLMToxicityDetector failing?

[rabah-khalek]

@rabah-khalek why is the LLMToxicityDetector failing?

You know what, I was completely blind to this. Focused more on the new detectors. Let me investigate this. Thanks!

[mattbit]

Also, minor observation: could you use ruff + black to format the code? I think we standardized on that (@andreybavt has configured a git pre-commit hook that you can enable).

[andreybavt]

https://docs.giskard.ai/en/latest/contribute/dev-environment.html

[rabah-khalek]

@mattbit, the issue is fixed now. The problem was that I was trying to input the target into the dataset by default, even though it might not exist, which was causing this problem in the scan.

Running detector LLMToxicityDetector… 0 issues detected. (Took 0:00:10.042270)
Running detector HarmfulnessDetector… 1 issues detected. (Took 0:00:13.708245)
Running detector GenderStereotypeDetector… 0 issues detected. (Took 0:00:22.539881)
Running detector MinorityStereotypeDetector… 1 issues detected. (Took 0:00:44.672983)
Scan completed: 2 issues found. (Took 0:01:30.972032)

[mattbit]

Could you review the comments?
Also, could you run black + ruff on new files to fix the style issues?

[sonarqubecloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

65.7% Coverage
0.0% Duplication