Mønstre for selvkritik og revision

COMPLETENESS_CRITIQUE = (
    'Original question: {question}\n\n'
    'Your previous answer: {answer}\n\n'
    'Check if your answer fully addresses the question:\n'
    '1. List each part or sub-question in the original question.\n'
    '2. For each part, indicate whether your answer addressed it.\n'
    '3. If any parts were missed or incomplete, provide a revised answer '
    'that covers everything.'
)

def check_completeness(question, initial_answer, client):
    prompt = COMPLETENESS_CRITIQUE.format(
        question=question,
        answer=initial_answer
    )
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=600,
        messages=[{'role': 'user', 'content': prompt}]
    )
    return response.content[0].text

HARM_CRITIQUE_PROMPT = (
    'Review the following assistant response for potential harms:\n\n'
    'User message: {user_message}\n'
    'Assistant response: {response}\n\n'
    'Consider:\n'
    '- Could this response enable harmful actions?\n'
    '- Could it be misused by someone with bad intentions?\n'
    '- Does it respect the privacy and dignity of individuals?\n'
    '- Could it cause psychological harm to vulnerable users?\n\n'
    'If the response has issues, explain them and provide a revised '
    'version that addresses the concerns while still being helpful. '
    'If the response is fine, say "Response is appropriate" and quote it back.'
)

def safety_check(user_message, response, client):
    prompt = HARM_CRITIQUE_PROMPT.format(
        user_message=user_message,
        response=response
    )
    result = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=600,
        messages=[{'role': 'user', 'content': prompt}]
    )
    return result.content[0].text

MULTI_CRITERIA_CRITIQUE = (
    'Evaluate this response on three criteria:\n\n'
    'Question: {question}\n'
    'Response: {response}\n\n'
    'Criteria:\n'
    '1. ACCURACY: Are all facts correct and claims well-supported?\n'
    '2. COMPLETENESS: Does it fully address the question?\n'
    '3. CLARITY: Is it easy to understand for the target audience?\n\n'
    'For each criterion, rate it Good/Fair/Poor and explain why.\n'
    'Then provide an improved response that scores Good on all three.'
)

def multi_criteria_check(question, response, client):
    prompt = MULTI_CRITERIA_CRITIQUE.format(
        question=question,
        response=response
    )
    result = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=800,
        messages=[{'role': 'user', 'content': prompt}]
    )
    return result.content[0].text

# Good revision prompt: action-oriented
GOOD_REVISION = (
    'Given this critique of your response, please write an improved version.\n\n'
    'Original question: {question}\n'
    'Critique: {critique}\n\n'
    'Write only the improved response — no preamble, no meta-commentary:'
)

# Bad revision prompt: too passive
BAD_REVISION = (
    'Here is a critique of your response. Can you maybe consider '
    'revising it somewhat based on the feedback below?\n'
    'Critique: {critique}'
    # Missing original question context!
    # Wishy-washy language reduces revision quality
)

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-...')

def multi_round_revision(question, n_rounds=2):
    # Round 0: Initial generation
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=512,
        messages=[{'role': 'user', 'content': question}]
    ).content[0].text

    principles = [
        'factual accuracy and avoiding unsupported claims',
        'completeness — ensuring all parts of the question are answered',
    ]

    for i, principle in enumerate(principles[:n_rounds]):
        critique_prompt = (
            f'Question: {question}\n'
            f'Current response: {response}\n\n'
            f'Critique for {principle} and provide an improved version:'
        )
        response = client.messages.create(
            model='claude-opus-4-5',
            max_tokens=512,
            messages=[{'role': 'user', 'content': critique_prompt}]
        ).content[0].text
        print(f'Round {i+1} complete')

    return response

CODE_CRITIQUE_PROMPT = (
    'Review this Python code for correctness and security issues:\n\n'
    'Task: {task}\n\n'
    'Code:\n'
    ''''python\n'
    '{code}\n'
    ''''\n\n'
    'Check for:\n'
    '1. Logic errors or off-by-one mistakes\n'
    '2. Unhandled edge cases (empty input, None, division by zero)\n'
    '3. Security issues (SQL injection, path traversal, etc.)\n\n'
    'If issues exist, list them and provide a corrected version. '
    'If the code is correct, say so and explain why it handles edge cases properly.'
)

def critique_code(task, code, client):
    prompt = CODE_CRITIQUE_PROMPT.format(task=task, code=code)
    result = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=800,
        messages=[{'role': 'user', 'content': prompt}]
    )
    return result.content[0].text

import anthropic
from collections import Counter

client = anthropic.Anthropic(api_key='sk-ant-...')

def self_consistency(question, n=5):
    """Generate N responses and pick the most common answer."""
    responses = []
    for _ in range(n):
        r = client.messages.create(
            model='claude-haiku-4-5',
            max_tokens=100,
            temperature=0.7,
            messages=[{'role': 'user', 'content': question}]
        )
        responses.append(r.content[0].text.strip())

    # Pick most common answer
    vote = Counter(responses)
    winner, count = vote.most_common(1)[0]
    print(f'Consensus ({count}/{n}): {winner}')
    return winner

result = self_consistency('What is the sum of angles in a triangle?')

def measure_critique_lift(examples, generate_fn, critique_fn, metric_fn):
    """
    Measure how much critique improves accuracy over initial generation.
    """
    initial_scores = []
    revised_scores = []

    for ex in examples:
        initial = generate_fn(ex.question)
        revised = critique_fn(ex.question, initial)

        initial_scores.append(metric_fn(ex.answer, initial))
        revised_scores.append(metric_fn(ex.answer, revised))

    avg_initial = sum(initial_scores) / len(initial_scores)
    avg_revised = sum(revised_scores) / len(revised_scores)

    print(f'Initial: {avg_initial:.1%}')
    print(f'Revised: {avg_revised:.1%}')
    print(f'Lift:   +{avg_revised - avg_initial:.1%}')
    return avg_revised - avg_initial