Skip to content

Majority vote give wrong label due to model inconsistency. #106

@Panzy-18

Description

@Panzy-18

In example/rater/generated_answer.ipynb. For a input of which true label is equivalent, model sometimes generate accept or reject. So majority vote can give wrong vote.

Input:

("Vitamin C (also known as ascorbic acid and ascorbate) is a water-soluble vitamin found in citrus and other fruits, berries and vegetables, also sold as a dietary supplement and as a topical serum ingredient to treat melasma (dark pigment spots) and wrinkles on the face.",
"Is Vitamin C water-soluble?",
"Yes, Vitamin C is a very water-soluble vitamin.",
"Yes, Vitamin C can be dissolved in water well."), # Equally good

Run:

config2 = RaterForGeneratedAnswerOpenAIGPT3p5Config()
config2.model_config.num_call = 3
config2.model_config.temperature = 0.9

with OpScope(name="TextFlow"):
    client2 = RaterClient(config2)

output = client2.run(data)
pprint.pprint(output)

Ouput:

{'output': [{'average_score': 0.0,
              'error': 'No errors.',
              'majority_vote': 'reject',
              'response': ['explanation: The grounding answer is better '
                           'because it directly states that Vitamin C is "very '
                           'water-soluble," while the generated answer is more '
                           'vague in saying that it "can be dissolved in water '
                           'well."\n'
                           'label: reject',
                           'explanation: Both the grounding answer and the '
                           'generated answer correctly state that Vitamin C is '
                           'water-soluble, so they are equivalent.\n'
                           'label: equivalent',
                           'explanation: The generated answer is better '
                           'because it accurately states that Vitamin C is '
                           'water-soluble, which aligns with the information '
                           'provided in the context.\n'
                           'label: accept'],
              'scores': [-1.0, 0.0, 1.0],
              'votes': ['reject', 'equivalent', 'accept']}],

Here 'majority_vote': 'reject' is wrong.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions