llm-evals/llmeval/commons/management/commands/eval_qa.py

import json
import re
import os
import sys
import time
import ollama
import dotenv
from openai import OpenAI
import google.generativeai as genai
from google.generativeai.types import HarmCategory, HarmBlockThreshold
import anthropic
from tqdm import tqdm
from abc import ABC, abstractmethod
from django.core.management.base import (
    BaseCommand, 
    CommandError, 
    CommandParser
)
from commons.models import (
    QA, Dataset, EvalAnswer,
    EvalConfig, RoleMessage, EvalSession
)


dotenv.load_dotenv()


class LLMClient(ABC):
    def __init__(self, model, session=None, **kwargs):
        self.messages = []
        self.model = model
        self.session = session
        self.final_answer_pattern = None
        self.stats = {
            "instruction": "",
            "answer": ""
        }

        if session and isinstance(session, EvalSession):
            self.final_answer_pattern = re.compile(session.config.final_answer_pattern)
            self.messages.append({
                "role": "system",
                "content": self.session.config.sys_prompt
            })

            for role_message in  RoleMessage.objects.filter(eval_config=session.config):
                self.messages.append({
                    "role": role_message.role,
                    "content": role_message.content
                })

    def make_messages(self, question, options, context):
        messages = self.messages.copy()
        option_str = self.option_str(options)

        content = f"Question: {question}"
        if option_str:
            content = f"{content}\nChoices:\n {option_str}"
        if context:
            content = f"{content}\nContext:\n{context}"

        messages.append(
            {"role": "user", 
             "content": content
             })
        return messages
    
    def get_chat_params(self):
        model_parameters = self.session.llm_model.parameters
        if self.session.parameters:
            model_parameters.update(self.session.parameters)
        return model_parameters

    @abstractmethod
    def send_question(self, question, options, context):
        pass
        
    def llm_eval(self, q):
        question = q.question
        options = q.options
        context = q.context
        correct_answer_idx = q.correct_answer_idx
        result = self.send_question(question, options, context)
        self.stats['answer'] = result
        match = re.search(self.final_answer_pattern, result)
        if not match:
            if self.session.answer_interpreter:
                interpreter_client = get_client(self.session.answer_interpreter.llm_model, self.session)
                question = self.session.answer_interpreter.prompt.replace("$QUESTION", self.stats['answer'])
                result = interpreter_client.send_question(question)
                self.stats['answer'] += f"\nInterpreter: {result}"
                match = re.search(self.final_answer_pattern, result)
        if match:
            final_answer = match.group(1)
            if final_answer.upper() == correct_answer_idx.upper():
                return True
        return False

    def messages_2_instruction(self, messages):
        instruction = ""
        for message in messages:
            instruction += f"{message['role'].upper()}: {message['content']}\n\n"
        return instruction
    
    def option_str(self, options=[]):
        options_str = ""
        if not options:
            return None
        for i, option in options.items():
            options_str += f"{i}) {option}\n"
        return options_str

class GoogleGenAI(LLMClient):
    def __init__(self, model, session, **kwargs):
        super().__init__(model, session, **kwargs)
        if 'api_key' not in kwargs:
            raise CommandError('Google Gen AI API key not found')
        genai.configure(api_key=kwargs['api_key'])
        self.client = genai.GenerativeModel(
            model_name = model, 
            system_instruction=self.session.config.sys_prompt
            )

    def send_question(self, question, options=[], context=None):
        messages = self.make_messages(question, options, context)
        self.stats['instruction'] = self.messages_2_instruction(messages)

        prompt = messages[-1]['content']
        messages = messages[:-1]

        history = []
        for message in messages:
            if message['role'] == 'system':
                continue
            if message['role'] == 'assistant':
                history.append({
                    "role": "model",
                    "parts": message['content']
                })
                continue
            history.append({
                "role": message['role'],
                "parts": message['content']
            })
        chat = self.client.start_chat(history=history)
        response = chat.send_message(prompt, safety_settings={
            HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_NONE,
            HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
            HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE,
            HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE,
        })

        return response.text
    
class OpenAIClient(LLMClient):
    def __init__(self, model, session, **kwargs):
        super().__init__(model, session, **kwargs)
        self.client = OpenAI(**kwargs)

    def send_question(self, question, options=[], context=None):
        messages = self.make_messages(question, options, context)
        self.stats['instruction'] = self.messages_2_instruction(messages)

        model_parameters = self.get_chat_params()

        completion = self.client.chat.completions.create(
            model=self.model,
            messages=messages,
            **model_parameters
        )
        msg_content = completion.choices[0].message.content
        return msg_content
    
class AnthropicClient(LLMClient):
    def __init__(self, model, session, **kwargs):
        super().__init__(model, session, **kwargs)
        if 'api_key' not in kwargs:
            raise CommandError('Anthropic API key not found')
        self.client = anthropic.Anthropic(api_key=kwargs['api_key'])

    def send_question(self, question, options=[], context=None):
        messages = self.make_messages(question, options, context)
        self.stats['instruction'] = self.messages_2_instruction(messages)

        sys_prompt = messages[0]['content']
        messages = messages[1:]

        # iterate the messages and concatenate the successive messages that have the same role
        new_messages = []
        for message in messages:
            if new_messages and new_messages[-1]['role'] == message['role']:
                new_messages[-1]['content'] += "\n" + message['content']
            else:
                new_messages.append(message)

        model_parameters =self.get_chat_params()
        
        response = self.client.messages.create(
            model=self.model,
            system=sys_prompt,
            messages=new_messages,
            **model_parameters
        )
        return response.content[0].text

class OllamaClient(LLMClient):
    def __init__(self, model, session, **kwargs):
        super().__init__(model, session, **kwargs)
        self.client = ollama.Client(**kwargs)

    def send_question(self, question, options, context=None):
        messages = self.make_messages(question, options, context)
        self.stats['instruction'] = self.messages_2_instruction(messages)

        model_parameters = self.get_chat_params()
        response = self.client.chat(model=self.model, messages=messages, **model_parameters)

        self.stats['answer'] = response['message']['content']
        return response['message']['content']

def get_client(llm_model, session):
    llm_backend = llm_model.backend
    if llm_backend.client_type == 'openai':
        if not llm_backend.parameters:
            raise CommandError('OpenAI parameters not found')
        return OpenAIClient(llm_model.name, session, **llm_backend.parameters)
    elif llm_backend.client_type == 'ollama':
        if not llm_backend.parameters:
            raise CommandError('Ollama parameters not found')
        return OllamaClient(llm_model.name, session, **llm_backend.parameters)
    elif llm_backend.client_type == 'genai':
        if not llm_backend.parameters:
            raise CommandError('Google GenAI parameters not found')
        return GoogleGenAI(llm_model.name, session, **llm_backend.parameters)
    elif llm_backend.client_type == 'anthropic':
        if not llm_backend.parameters:
            raise CommandError('Anthropic parameters not found')
        return AnthropicClient(llm_model.name, session, **llm_backend.parameters)

class Command(BaseCommand):
    help = 'Evaluate MedQA'

    def add_arguments(self, parser: CommandParser) -> None:
        parser.add_argument('--sample-size', type=int, help='Sample size')
        parser.add_argument('--randomize', action='store_true', help='Randomize questions')
        parser.add_argument('--session-id', type=int, help='Session ID')
        parser.add_argument('--continue', action='store_true', help='Continue from last question')

    def handle(self, *args, **options):
        sample_size = options['sample_size']
        randomize = options['randomize']
        session_id = options['session_id']
        continue_from_last = options['continue']
        
        session = EvalSession.objects.filter(id=session_id).first()
        if not session:
            raise CommandError('Session not found')
        
        dataset = session.config.dataset
        target = session.dataset_target
        llm_model = session.llm_model
        llm_backend = llm_model.backend
        

        qa_set = QA.objects.filter(dataset=dataset).order_by('id')
        
        if target:
            qa_set = qa_set.filter(target=target)

        if randomize:
            qa_set = qa_set.order_by('?')

        if sample_size:
            qa_set = qa_set[:sample_size]

        if not qa_set.exists():
            raise CommandError('No questions found')
        
        if continue_from_last:
            for q in qa_set:
                if EvalAnswer.objects.filter(eval_session=session, question=q).exists():
                    qa_set = qa_set.exclude(id=q.id)

        print(f"Questions to evaluate: {qa_set.count()}")

        client = get_client(llm_model, session)
        
        stats = {
            "correct": 0,
            "total": qa_set.count()
        }
        for q in tqdm(qa_set, desc="Evaluating dataset"):
            correct = client.llm_eval(q)
            eval_answer = EvalAnswer(
                eval_session=session,
                question=q,
                is_correct=correct,
                instruction=client.stats['instruction'],
                assistant_answer = client.stats['answer'],
                hash=q.hash,
                llm_backend=llm_backend,
                llm_model=llm_model
            )
            if correct:
                stats["correct"] += 1
            
            eval_answer.save()
            if session.request_delay:
                time.sleep(session.request_delay)

        print(f"Accuracy: {stats['correct']}/{stats['total']} ({stats['correct']/stats['total']:.2f})")

        self.stdout.write(self.style.SUCCESS('Successfully evaluated MedQA'))