| 6 | |
| 7 | |
| 8 | class Claude3Evaluator(Evaluator): |
| 9 | |
| 10 | def __init__(self, pretrained_model_name_or_path, api_key, temperature=0): |
| 11 | super(Claude3Evaluator, self).__init__() |
| 12 | |
| 13 | self.client = anthropic.Anthropic( |
| 14 | api_key=api_key |
| 15 | ) |
| 16 | self.model = pretrained_model_name_or_path |
| 17 | self.temperature = temperature |
| 18 | self.max_tokens = 4096 |
| 19 | |
| 20 | def format_prompt(self, prompt): |
| 21 | return [ |
| 22 | { |
| 23 | 'role': 'user', |
| 24 | 'content': prompt |
| 25 | } |
| 26 | ] |
| 27 | |
| 28 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 29 | def generate_text(self, prompt): |
| 30 | prompt = self.format_prompt(prompt) |
| 31 | response = self.client.messages.create( |
| 32 | model=self.model, |
| 33 | messages=prompt, |
| 34 | temperature=self.temperature, |
| 35 | max_tokens=self.max_tokens |
| 36 | ) |
| 37 | |
| 38 | return response.content[0].text.strip() |
| 39 | |
| 40 | @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6)) |
| 41 | def count_tokens(self, prompt): |
| 42 | return self.client.count_tokens(prompt) |