Evaluating Base and Finetuned LLMs

In this tutorial, we will evaluate the performance of a base and finetuned LLM while comparing it to the performance of OpenAI. For this example, the base Llama2-7B and finetuned Llama2-7B models will be used.

Prerequisites

• You need to ingest your data corpus and create a dataset from it. You can refer to the Data ingestion and creating dataset tutorial for more information on how to do this.

  • The dataset name used in this tutorial is contracts.

• You need to finetune the Llama2-7b model on DKubeX. For a comprehensive guide on how to finetune a model, refer to the Finetuning Open Source LLMs tutorial.

• You need to deploy the base Llama2-7B and the finetuned Llama2-7B model on DKubeX. To learn how to deploy a LLM on DKubeX, refer to the Deploying LLMs in DKubeX tutorial.

  • The names of the base and finetuned Llama2-7B deployments used in this tutorial are llama27bbase and llama27bft respectively.

• Export the following variables to your workspace by running the following commands on your DKubeX Terminal.

  • Replace the <username> part with your DKubeX workspace name.

    export NAMESPACE="<username>"
    export HOMEDIR=/home/${NAMESPACE}
    

• A few .yaml files are required to be used in the evaluation process.

  • On the Terminal application in DKubeX UI, run the following commands:

    git clone -b v0.8.3 https://github.com/dkubeio/dkubex-examples.git
    cd && cp dkubex-examples/rag/query/query.yaml ${HOMEDIR}/query.yaml && cp dkubex-examples/rag/evaluation/eval.yaml ${HOMEDIR}/eval.yaml && cd
    

Evaluating LLM Models

In this example, first we will evaluate the base Llama2-7B model comparing it to the performance of OpenAI, and then follow the same steps to evaluate the finetuned Llama2-7B model.

To evaluate the base Llama2-7B model, follow the steps provided below:

• Provide the appropriate details on the query.yaml file which will be used during the evaluation process. Run vim query.yaml and provide the following details:

  Note

  On the chat_engine:url: section, provide the endpoint URL of the deployed model to be used. The endpoint URL can be found on the Deployments page of DKubeX UI.

  query.yaml DescriptionExample

  Field	Sub-field	Description
  input	question	The input question to be answered by the RAG system.
  	mode	The mode of interaction with the pipeline.
  vectorstore_retriever	kind	Specifies the type of vector store retriever.
  	provider	Provider for the vector store retriever.
  	embedding_class	Class of embedding used for retrieval.
  	embedding_model	Name of the embedding model from HuggingFace.
  	dataset	Name of the ingested dataset.
  	textkey	Key identifying the text data within the dataset.
  	top_k	The number of results to retrieve per query.
  prompt_builder	prompt_str	The prompt string used for generation.
  	prompt_file	The file containing the prompt string.
  nodes_sorter	max_sources	Maximum number of sources to consider during sorting.
  reranker	model	Name of the re-ranker model from Hugging Face.
  	top_n	The number of results to re-rank.
  contexts_joiner	separator	Separator used for joining different contexts.
  chat_engine	llm	Specifies the LLM to be used for generation.
  	url	Service URL for the LLM deployment to be used.
  	llmkey	Authentication key for accessing the LLM service.
  	window_size	Size of the window for context generation.
  	max_tokens	Maximum number of tokens for generation.
  tracking	experiment	MLflow experiment name for tracking.

• Provide the appropriate details on the eval.yaml file which will be used during the evaluation process. Run vim eval.yaml and provide the following details:

  Note

  • Provide your own OpenAI API key on the questions_generator:llmkey: section.

  • On the semantic_similarity_evaluator:llmurl: section, provide the endpoint URL of the deployed model to be used. The endpoint URL can be found on the Deployments page of DKubeX UI.

  • On the semantic_similarity_evaluator:llmkey: section, provide the serving token of the deployed model to be used. To find the serving token, go to the Deployments page of DKubeX UI, click on the deployed model, and copy the serving token from the Serving Token section.

  eval.yaml DescriptionExample

  Field	Sub-field		Description
  vectorstore_reader	kind		Specifies the type of vectorstore reader.
  	provider		Indicates the provider of the vectorstore reader.
  	properties		Lists the properties of the vectorstore reader.
  questions_generator	prompt_str		Defines the strategy for generating prompts.
  	prompt_file		File containing custom prompt.
  	num_questions_per_chunk		Specifies the number of questions to generate per data chunk.
  	max_chunks		Sets the maximum number of data chunks to generate questions.
  	llm		Determines the language model (LLM) to use for generating questions.
  	llm_key		Requires providing the API key for the chosen LLM.
  	llmurl		Indicates the URL where the chosen LLM service is deployed.
  	max_tokens		Specifies the maximum number of tokens allowed in each question.
  retrieval_evaluator	vector_retriever	kind	Indicates the type of vector retriever.
  		provider	Specifies the provider of the vector retriever.
  		textkey	Refers to the key used to access the text data within the vector retriever.
  		embedding_model	Specifies the name of the embedding model used for text representation.
  		similarity_top_k	Sets the number of similar items to retrieve for each query.
  	metrics		Specifies the evaluation metrics used for retrieval evaluation.
  semantic_similarity_evaluator	prompt_str		Defines the strategy for similarity evaluation.
  	prompt_file		File containing custom prompt.
  	llm		Specifies the language model (LLM) to use for semantic similarity evaluation.
  	llmkey		Labeled dummy in case of local deployments available within DKubeX or used to pass auth key if using an externam endpoint.
  	llmurl		Indicates the URL where the chosen LLM service is deployed.
  	max_tokens		Specifies the maximum number of tokens allowed in each semantic similarity evaluation prompt.
  	metrics		Specifies the evaluation metric used for semantic similarity evaluation.
  tracking	experiment		Provides a unique name for the MLFlow experiment, allowing for tracking and comparison of different runs of the pipeline.

• Once done, run the following command to start the evaluation process of the base Llama2-7B model. Replace the <dataset name> part with the name of the dataset created during ingestion (for this example, contracts).

  CommandExample

  d3x dataset evaluate -d <dataset name> --config ${HOMEDIR}/eval.yaml