tutorial.md

Evaluate Tutorial

🚀 How to Evaluate

If you need to test your own huggingface-formatted model, the overall steps are as follows:

1. Write the loader function for the model.
2. Write the context_builder function for the model.
3. Register the model in the configuration file.
4. Run the testing script. If the model does not require any special processing after loading, and the input does not need to be converted to a specific format (e.g. chatml format or other human-bot formats), you can directly proceed to step 4 to initiate the testing.

1. Write the loader function

If the model requires additional processing after loading (e.g. adjusting the tokenizer), you need to inherit the ModelAndTokenizerLoader class in src.context_builder.context_builder_family.py and override the corresponding load_model and load_tokenizer functions. You can refer to the following example:

class QwenModelAndTokenizerLoader(ModelAndTokenizerLoader):
    def __init__(self):
      super().__init__()
      pass
    
    def load_model(self, model_path: str):
        model = super().load_model(model_path)
        model.generation_config = GenerationConfig.from_pretrained(model_path)
        return model

    def load_tokenizer(self, model_path: str):
        tokenizer = super().load_tokenizer(model_path)
    
        # read generation config
        with open(model_path + '/generation_config.json', 'r') as f:
        generation_config = json.load(f)
        tokenizer.pad_token_id = generation_config['pad_token_id']
        tokenizer.eos_token_id = generation_config['eos_token_id']
        return tokenizer

2. Write the context_builder function for the Model

If the input needs to be converted to a specific format (e.g. chatml format or other human-bot formats), you need to inherit the ContextBuilder class in src.context_builder.context_builder_family and override the make_context function. This function is used to convert the input to the corresponding required format. An example is shown below:

class QwenChatContextBuilder(ContextBuilder):
    def __init__(self):
        super().__init__()
    
    def make_context(
        self,
        model,
        tokenizer, 
        query: str,
        system: str = "you are a helpful assistant"
    ):
        '''
        model: PretrainedModel
        tokenizer: PretrainedTokenzier
        query: Input string
        system: System prompt if needed
        '''
        im_start, im_end = "<|im_start|>", "<|im_end|>"
        im_start_tokens = [tokenizer.im_start_id]
        im_end_tokens = [tokenizer.im_end_id]
        nl_tokens = tokenizer.encode("\n")

        def _tokenize_str(role, content):
            return f"{role}\n{content}", tokenizer.encode(
                role, allowed_special=set()
            ) + nl_tokens + tokenizer.encode(content, allowed_special=set())

        system_text, system_tokens_part = _tokenize_str("system", system)
        system_tokens = im_start_tokens + system_tokens_part + im_end_tokens

        raw_text = ""
        context_tokens = []

        context_tokens = system_tokens + context_tokens
        raw_text = f"{im_start}{system_text}{im_end}" + raw_text
        context_tokens += (
            nl_tokens
            + im_start_tokens
            + _tokenize_str("user", query)[1]
            + im_end_tokens
            + nl_tokens
            + im_start_tokens
            + tokenizer.encode("assistant")
            + nl_tokens
        )
        raw_text += f"\n{im_start}user\n{query}{im_end}\n{im_start}assistant\n"
        return raw_text, context_tokens

3. Register the model in the configuration file

Go to the model_conf.json file in the conf directory and register the corresponding model name and the loader and context_builder that will be used for this model. Simply write the class names defined in the first and second steps for the loader and context_builder. Here is an example:

{
  "Qwen-Chat": {
  "loader": "QwenModelAndTokenizerLoader",
  "context_builder": "QwenChatContextBuilder"
  }
}

4. Execute the testing script

Run the following code to initiate the test:

# model_path: path to the model for testing
# model_name: the model name corresponding to the model in the configuration file, default is Default, which represents using the default loader and context_builder
# model_conf_path: path to the model configuration file, usually the devopseval_dataset_fp.json file in the conf directory
# eval_dataset_list: the names of the datasets to be tested, default is all to test all datasets, if you need to test one or more datasets, use the # symbol to connect them, for example: dataset1#dataset2
# eval_dataset_fp_conf_path: path to the dataset configuration file
# eval_dataset_type: the type of testing, only supports the default test type of test dataset
# data_path: path to the evaluation dataset, fill in the downloaded dataset address
# k_shot: supports 0-5, represents the number of example prefixes added for few-shot

python src/run_eval.py \
--model_path path_to_model \
--model_name model_name_in_conf \
--model_conf_path path_to_model_conf \
--eval_dataset_list all \
--eval_dataset_fp_conf_path path_to_dataset_conf \
--eval_dataset_type test \
--data_path path_to_downloaded_devops_eval_data \
--k_shot 0

For example, if the evaluation dataset is downloaded to folder1, the code is placed in folder2, and the model is in folder3, and the model does not require custom loader and context_builder, and all zero-shot scores of all datasets need to be tested, you can use the following script to initiate the test:

python folder2/src/run_eval.py \
--model_path folder3 \
--model_name Default \
--model_conf_path folder1/conf/model_conf.json \
--eval_dataset_list all \
--eval_dataset_fp_conf_path folder1/conf/devopseval_dataset_fp.json \
--eval_dataset_type test \
--data_path folder2 \
--k_shot 0