Add Doc Strings to Config Files

examples/model_configs/peft_model.yaml

@@ -1,6 +1,6 @@
 model:
   base_params:
-    model_args: "pretrained=predibase/customer_support,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ... For a PEFT model, the pretrained model should be the one trained with PEFT and the base model below will contain the original model on which the adapters will be applied.
+    model_args: "pretrained=predibase/customer_support,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ... For a PEFT model, the pretrained model should be the one trained with PEFT and the base model below will contain the original model on which the adapters will be applied. To see the full list of parameters, please see here: https://huggingface.co/docs/lighteval/main/en/package_reference/models#lighteval.models.transformers.adapter_model.AdapterModelConfig
     dtype: "4bit"  # Specifying the model to be loaded in 4 bit uses BitsAndBytesConfig. The other option is to use "8bit" quantization.
     compile: true
   merged_weights: # Ignore this section if you are not using PEFT models

examples/model_configs/quantized_model.yaml

@@ -1,9 +1,9 @@
 model:
   base_params:
-    model_args: "pretrained=HuggingFaceH4/zephyr-7b-beta,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ...
+    model_args: "pretrained=HuggingFaceH4/zephyr-7b-beta,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True.  To see the full list of parameters, please see here: https://huggingface.co/docs/lighteval/main/en/quicktour#model-arguments .
     dtype: "4bit"  # Specifying the model to be loaded in 4 bit uses BitsAndBytesConfig. The other option is to use "8bit" quantization.
     compile: true
-  merged_weights: # Ignore this section if you are not using PEFT models
+  merged_weights: # Ignore this section if you are not using PEFT models.
     delta_weights: false # set to True of your model should be merged with a base model, also need to provide the base model name
     adapter_weights: false # set to True of your model has been trained with peft, also need to provide the base model name
     base_model: null # path to the base_model - needs to be specified only if delta_weights or adapter_weights is set to True

examples/model_configs/serverless_model.yaml

@@ -1,3 +1,3 @@
 model:
   base_params:
-    model_name: "meta-llama/Llama-3.1-8B-Instruct" #Qwen/Qwen2.5-14B" #Qwen/Qwen2.5-7B"
+    model_name: "meta-llama/Llama-3.1-8B-Instruct" #Qwen/Qwen2.5-14B" #Qwen/Qwen2.5-7B".  #To see the full list of parameters, please see here: https://huggingface.co/docs/lighteval/package_reference/models#endpoints-based-models

examples/model_configs/tgi_model.yaml

@@ -2,4 +2,4 @@ model:
   instance:
     inference_server_address: ""
     inference_server_auth: null
-    model_id: null # Optional, only required if the TGI container was launched with model_id pointing to a local directory
+    model_id: null # Optional, only required if the TGI container was launched with model_id pointing to a local directory. To see the full list of parameters, please see here: https://huggingface.co/docs/lighteval/package_reference/models#lighteval.models.endpoints.tgi_model.TGIModelConfig

examples/model_configs/transformers_model.yaml

@@ -1,6 +1,6 @@
 model:
   base_params:
-    model_args: "pretrained=HuggingFaceTB/SmolLM-1.7B,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ...
+    model_args: "pretrained=HuggingFaceH4/zephyr-7b-beta,revision=main" # pretrained=model_name,trust_remote_code=boolean,revision=revision_to_use,model_parallel=True ...
     dtype: "bfloat16"
     compile: true
   merged_weights: # Ignore this section if you are not using PEFT models
@@ -9,4 +9,3 @@ model:
     base_model: null # path to the base_model
   generation:
     multichoice_continuations_start_space: null # If true/false, will force multiple choice continuations to start/not start with a space. If none, will do nothing
-    temperature: 0.5

src/lighteval/models/endpoints/endpoint_model.py

@@ -97,6 +97,28 @@ def from_path(cls, path: str) -> "ServerlessEndpointModelConfig":
 
 @dataclass
 class InferenceEndpointModelConfig:
+    """
+    This class defines the settings used to deploy inference endpoints automatically. (Inference endpoint docs: https://huggingface.co/docs/inference-endpoints/index)
+
+        Attributes:
+        endpoint_name (str, Optional): Inference endpoint name (to use at creation or reuse)
+        model_name (str, optional): The name of the model for inference.
+        reuse_existing (bool, default: False): Indicates whether to reuse an existing endpoint.
+        accelerator (str, default: "gpu"): Specifies the type of hardware accelerator.
+        model_dtype (str, optional): The data type used by the model. Defaults to the framework's choice if None.
+        vendor (str, default: "aws"): Cloud service provider for hosting the endpoint.
+        region (str, default: "us-east-1"): Cloud region, chosen based on hardware availability.
+        instance_size (str, optional): Specifies the size of the instance (e.g., large, xlarge).
+        instance_type (str, optional): Specifies the type of the instance (e.g., g5.4xlarge).
+        framework (str, default: "pytorch"): Framework used for inference (e.g., pytorch, tensorflow).
+        endpoint_type (str, default: "protected"): Security level of the endpoint (e.g., public, protected).
+        add_special_tokens (bool, default: True): Specifies if special tokens should be added during processing.
+        revision (str, default: "main"): The Git branch or commit hash of the model.
+        namespace (str, optional): The namespace under which the endpoint is launched.
+        image_url (str, optional): Docker image URL for the endpoint.
+        env_vars (dict, optional): Environment variables for the endpoint.
+    """
+
     endpoint_name: str = None
     model_name: str = None
     reuse_existing: bool = False

src/lighteval/models/endpoints/openai_model.py

@@ -62,15 +62,48 @@
 
 @dataclass
 class OpenAIModelConfig:
+    """
+    Configuration class to create an [[OpenAIModel]], to call via its API at inference for evaluation.
+
+    Attributes:
+    model (str): name or identifier of the OpenAI model to be used for inference.
+    generation_parameters(None,GenerationParameters): Parameters for model generation. If not
+                                                     provided, defaults to a new instance
+                                                     of `GenerationParameters`.
+    """
+
     model: str
     generation_parameters: GenerationParameters = None
 
     def __post_init__(self):
+        """
+
+        Post-initialization that ensures the `generation_parameters` is set
+        to a valid `GenerationParameters`. If not provided, initializes a default one.
+        """
         if not self.generation_parameters:
             self.generation_parameters = GenerationParameters()
 
     @classmethod
     def from_path(cls, path: str) -> "OpenAIModelConfig":
+        """
+        Creates an instance of `OpenAIModelConfig` from a YAML configuration file.
+
+        Loads the model configuration from a given file path and initializes the
+        `OpenAIModelConfig` with the model name and corresponding `GenerationParameters` parsed
+        from the file.
+
+        Args:
+            path (str): Path to the YAML configuration file containing the model configuration.
+
+        Returns:
+            OpenAIModelConfig: An instance of `OpenAIModelConfig` with the configuration loaded
+                               from the specified YAML file.
+
+        Raises:
+            FileNotFoundError: If the specified file path does not exist.
+            KeyError: If required keys are missing in the YAML configuration file.
+        """
         import yaml
 
         with open(path, "r") as f:
@@ -162,11 +195,11 @@ def greedy_until(
         Generates responses using a greedy decoding strategy until certain ending conditions are met.
 
         Args:
-            requests (list[Request]): list of requests containing the context and ending conditions.
+            requests (list[GreedyUntilRequest]): list of requests containing the context and ending conditions.
             override_bs (int, optional): Override the batch size for generation. Defaults to None.
 
         Returns:
-            list[GenerativeResponse]: list of generated responses.
+            list [GenerativeResponse]: list of generated responses.
         """
         for request in requests:
             request.tokenized_context = self.tok_encode(request.context)
@@ -251,7 +284,7 @@ def _loglikelihood_tokens(
             ), "Only single token continuations are supported when using openai API."
 
             for i in range(len(dataset)):
-                logit_bias = {tok: 100 for tok in dataset[i].tokenized_continuation}
+                logit_bias = dict.fromkeys(dataset[i].tokenized_continuation, 100)
                 logit_biass.append(logit_bias)
 
             outputs = self.__call_api_parallel(

src/lighteval/models/endpoints/tgi_model.py

@@ -48,6 +48,15 @@ def divide_chunks(array, n):
 
 @dataclass
 class TGIModelConfig:
+    """
+    Provides a streamlined configuration for integrating with Text Generation Inference (TGI) endpoints. To know more, please click here: https://huggingface.co/docs/text-generation-inference/index
+
+    Attributes:
+    inference_server_address (str, required):Endpoint address of the inference server hosting the model.
+    inference_server_auth (str, required): Authentication credentials or tokens required to access the server.
+    model_id (str, required): Identifier for the model hosted on the inference server.
+    """
+
     inference_server_address: str
     inference_server_auth: str
     model_id: str

src/lighteval/models/transformers/adapter_model.py

@@ -27,7 +27,7 @@
 import torch
 from transformers import AutoModelForCausalLM, PreTrainedTokenizer
 
-from lighteval.models.transformers.transformers_model import TransformersModel, TransformersModelConfig
+from lighteval.models.transformers.base_model import BaseModel, BaseModelConfig
 from lighteval.models.utils import _get_dtype
 from lighteval.utils.imports import NO_PEFT_ERROR_MSG, is_peft_available
 from lighteval.utils.utils import EnvConfig
@@ -40,7 +40,16 @@
 
 
 @dataclass
-class AdapterModelConfig(TransformersModelConfig):
+class AdapterModelConfig(BaseModelConfig):
+    """
+    Manages the configuration of adapter models. Adapter models are designed to extend or adapt a
+    base model's functionality for specific tasks while keeping most of the base model's parameters frozen.
+
+    Attributes:
+        base_model (str): The name of the parent base model. This model provides the tokenizer and configuration for the adapter model.
+                          Defaults to None if not specified.
+    """
+
     # Adapter models have the specificity that they look at the base model (= the parent) for the tokenizer and config
     base_model: str = None
 
@@ -54,11 +63,33 @@ def __post_init__(self):
         return super().__post_init__()
 
     def init_configs(self, env_config: EnvConfig):
+        """
+        Initializes the configurations of adapter models.
+
+        Args:
+            env_configs(EnvConfig): An instance of EnvConfig.
+
+        Returns:
+            Any: Result of the configuration initialization.
+        """
         return self._init_configs(self.base_model, env_config)
 
 
-class AdapterModel(TransformersModel):
+class AdapterModel(BaseModel):
+    """
+    Integrates the adapter models with a pre-trained base model.
+    """
+
     def _create_auto_tokenizer(self, config: AdapterModelConfig, env_config: EnvConfig) -> PreTrainedTokenizer:
+        """
+        Creates and configures the adapter model by applying adapter weights to the base model.
+
+        Args:
+            config(AdapterModelConfig): An instance of AdapterModelConfig.
+            env_config(EnvConfig): An instance of EnvConfig.
+
+        Returns: PreTrainedTokenizer
+        """
         # By default, we look at the model config for the model stored in `base_model`
         # (= the parent model, not the model of interest)
         return self._create_auto_tokenizer_with_name(
@@ -71,7 +102,15 @@ def _create_auto_tokenizer(self, config: AdapterModelConfig, env_config: EnvConf
         )
 
     def _create_auto_model(self, config: AdapterModelConfig, env_config: EnvConfig) -> AutoModelForCausalLM:
-        """Returns a PeftModel from a base model and a version fined tuned using PEFT."""
+        """
+        Returns a PeftModel from a base model and a version fined tuned using PEFT.
+
+        Args:
+            config(AdapterModelConfig): An instance of AdapterModelConfig.
+            env_config(EnvConfig): An instance of EnvConfig.
+
+        Returns: AutoModelForCasualLM
+        """
         torch_dtype = _get_dtype(config.dtype, self._config)
         config.model_parallel, max_memory, device_map = self.init_model_parallel(config.model_parallel)
 

src/lighteval/models/transformers/delta_model.py

@@ -28,7 +28,7 @@
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM
 
-from lighteval.models.transformers.transformers_model import TransformersModel, TransformersModelConfig
+from lighteval.models.transformers.base_model import BaseModel, BaseModelConfig
 from lighteval.models.utils import _get_dtype, _get_model_sha
 from lighteval.utils.utils import EnvConfig
 
@@ -37,7 +37,11 @@
 
 
 @dataclass
-class DeltaModelConfig(TransformersModelConfig):
+class DeltaModelConfig(BaseModelConfig):
+    """
+    This class is used to manage the configuration class for delta models.
+    """
+
     # Delta models look at the pretrained (= the delta weights) for the tokenizer and model config
     base_model: str = None
 
@@ -53,13 +57,21 @@ def get_model_sha(self):
         return _get_model_sha(repo_id=self.pretrained, revision="main")
 
 
-class DeltaModel(TransformersModel):
+class DeltaModel(BaseModel):
     def _create_auto_model(
         self,
         config: DeltaModelConfig,
         env_config: EnvConfig,
     ) -> AutoModelForCausalLM:
-        """Returns a model created by adding the weights of a delta model to a base model."""
+        """
+        It returns a model created by adding the weights of a delta model to a base model.
+
+         Args:
+        config(AdapterModelConfig): An instance of AdapterModelConfig.
+        env_config(EnvConfig): An instance of EnvConfig.
+
+        Returns: AutoModelForCasualLM
+        """
         config.model_parallel, max_memory, device_map = self.init_model_parallel(config.model_parallel)
         torch_dtype = _get_dtype(config.dtype, self._config)
 

src/lighteval/models/vllm/vllm_model.py

@@ -75,6 +75,29 @@
 
 @dataclass
 class VLLMModelConfig:
+    """
+    This class defines the configuration parameters for deploying and running models using the vLLM framework.
+
+    Attributes:
+    pretrained (str, required): The identifier for the pretrained model (e.g., model name or path).
+    gpu_memory_utilisation (float, default: 0.9): Fraction of GPU memory to allocate for the model. Reduce this value if you encounter memory issues.
+    revision (str, default: "main"): Specifies the branch or version of the model repository.
+    dtype (str | None, optional): Data type for computations (e.g., float32, float16, or bfloat16). Defaults to the model's preset if None.
+    tensor_parallel_size (int, default: 1): Number of GPUs used for splitting tensors across devices.
+    pipeline_parallel_size (int, default: 1): Number of GPUs used for pipeline parallelism.
+    data_parallel_size (int, default: 1): Number of GPUs used for data parallelism.
+    max_model_length (int | None, optional): Maximum sequence length for the model. If None, it is inferred automatically. Can be reduced to handle Out-of-Memory (OOM) issues.
+    swap_space (int, default: 4): Amount of CPU swap space (in GiB) per GPU for offloading.
+    seed (int, default: 1234): Seed for reproducibility in experiments.
+    trust_remote_code (bool, default: False): Whether to trust custom code provided by remote repositories.
+    use_chat_template (bool, default: False): Specifies if chat-specific templates should be used for input formatting.
+    add_special_tokens (bool, default: True): Indicates whether to add special tokens during tokenization.
+    multichoice_continuations_start_space (bool, default: True): Adds a space at the beginning of each continuation during multi-choice generation.
+    pairwise_tokenization (bool, default: False): Specifies if context and continuation should be tokenized separately or together.
+    subfolder (Optional[str], optional): Path to a specific subfolder in the model repository, if applicable.
+    temperature (float, default: 0.6): Sampling temperature for stochastic tasks. Ignored for deterministic tasks (set internally to 0).
+    """
+
     pretrained: str
     gpu_memory_utilisation: float = 0.9  # lower this if you are running out of memory
     revision: str = "main"  # revision of the model