Prepare Models

To support the evaluation of new models in OpenCompass, there are several ways:

1. HuggingFace-based models

2. API-based models

3. Custom models

HuggingFace-based Models

In OpenCompass, we support constructing evaluation models directly from HuggingFace’s AutoModel.from_pretrained and AutoModelForCausalLM.from_pretrained interfaces. If the model to be evaluated follows the typical generation interface of HuggingFace models, there is no need to write code. You can simply specify the relevant configurations in the configuration file.

Here is an example configuration file for a HuggingFace-based model:

# Use `HuggingFace` to evaluate models supported by AutoModel.
# Use `HuggingFaceCausalLM` to evaluate models supported by AutoModelForCausalLM.
from opencompass.models import HuggingFaceCausalLM

models = [
    dict(
        type=HuggingFaceCausalLM,
        # Parameters for `HuggingFaceCausalLM` initialization.
        path='huggyllama/llama-7b',
        tokenizer_path='huggyllama/llama-7b',
        tokenizer_kwargs=dict(padding_side='left', truncation_side='left'),
        max_seq_len=2048,
        batch_padding=False,
        # Common parameters shared by various models, not specific to `HuggingFaceCausalLM` initialization.
        abbr='llama-7b',            # Model abbreviation used for result display.
        max_out_len=100,            # Maximum number of generated tokens.
        batch_size=16,              # The size of a batch during inference.
        run_cfg=dict(num_gpus=1),   # Run configuration to specify resource requirements.
    )
]

Explanation of some of the parameters:

• batch_padding=False: If set to False, each sample in a batch will be inferred individually. If set to True, a batch of samples will be padded and inferred together. For some models, such padding may lead to unexpected results. If the model being evaluated supports sample padding, you can set this parameter to True to speed up inference.

• padding_side='left': Perform padding on the left side. Not all models support padding, and padding on the right side may interfere with the model’s output.

• truncation_side='left': Perform truncation on the left side. The input prompt for evaluation usually consists of both the in-context examples prompt and the input prompt. If the right side of the input prompt is truncated, it may cause the input of the generation model to be inconsistent with the expected format. Therefore, if necessary, truncation should be performed on the left side.

During evaluation, OpenCompass will instantiate the evaluation model based on the type and the initialization parameters specified in the configuration file. Other parameters are used for inference, summarization, and other processes related to the model. For example, in the above configuration file, we will instantiate the model as follows during evaluation:

model = HuggingFaceCausalLM(
    path='huggyllama/llama-7b',
    tokenizer_path='huggyllama/llama-7b',
    tokenizer_kwargs=dict(padding_side='left', truncation_side='left'),
    max_seq_len=2048,
)

API-based Models

Currently, OpenCompass supports API-based model inference for the following:

• OpenAI (opencompass.models.OpenAI)

• ChatGLM (opencompass.models.ZhiPuAI)

• ABAB-Chat from MiniMax (opencompass.models.MiniMax)

• XunFei from XunFei (opencompass.models.XunFei)

Let’s take the OpenAI configuration file as an example to see how API-based models are used in the configuration file.

from opencompass.models import OpenAI

models = [
    dict(
        type=OpenAI,                             # Using the OpenAI model
        # Parameters for `OpenAI` initialization
        path='gpt-4',                            # Specify the model type
        key='YOUR_OPENAI_KEY',                   # OpenAI API Key
        max_seq_len=2048,                        # The max input number of tokens
        # Common parameters shared by various models, not specific to `OpenAI` initialization.
        abbr='GPT-4',                            # Model abbreviation used for result display.
        max_out_len=512,                         # Maximum number of generated tokens.
        batch_size=1,                            # The size of a batch during inference.
        run_cfg=dict(num_gpus=0),                # Resource requirements (no GPU needed)
    ),
]

We have provided several examples for API-based models. Please refer to

configs
├── eval_zhipu.py
├── eval_xunfei.py
└── eval_minimax.py

Custom Models

If the above methods do not support your model evaluation requirements, you can refer to Supporting New Models to add support for new models in OpenCompass.