llama-index-packs-llama-dataset-metadata 0.4.0

llama-index packs llama_dataset_metadata integration

LlamaDataset Metadata Pack

As part of the LlamaDataset submission package into llamahub, two metadata files are required, namely: card.json and README.md. This pack creates these two files and saves them to disk to help expedite the submission process.

CLI Usage

You can download llamapacks directly using llamaindex-cli, which comes installed with the llama-index python package:

llamaindex-cli download-llamapack LlamaDatasetMetadataPack --download-dir ./llama_dataset_metadata_pack

You can then inspect the files at ./llama_dataset_metadata_pack and use them as a template for your own project!

Code Usage

You can download the pack to the ./llama_dataset_metadata_pack directory through python code as well. The sample script below demonstrates how to construct LlamaDatasetMetadataPack using a LabelledRagDataset downloaded from llama-hub and a simple RAG pipeline built off of its source documents.

from llama_index.core.llama_pack import download_llama_pack

# Download and install dependencies
LlamaDatasetMetadataPack = download_llama_pack(
    "LlamaDatasetMetadataPack", "./llama_dataset_metadata_pack"
)

# construction requires a query_engine, a rag_dataset, and optionally a judge_llm
llama_dataset_metadata_pack = LlamaDatasetMetadataPack()

# create and save `card.json` and `README.md` to disk
dataset_description = (
    "A labelled RAG dataset based off an essay by Paul Graham, consisting of "
    "queries, reference answers, and reference contexts."
)

llama_dataset_metadata_pack.run(
    name="Paul Graham Essay Dataset",
    description=dataset_description,
    rag_dataset=rag_dataset,  # defined earlier not shown here
    index=index,  # defined earlier not shown here
    benchmark_df=benchmark_df,  # defined earlier not shown here
    baseline_name="llamaindex",
)

NOTE: this pack should be used only after performing a RAG evaluation (i.e., by using RagEvaluatorPack on a LabelledRagDataset). In the code snippet above, index, rag_dataset, and benchmark_df are all objects that you'd expect to have only after performing the RAG evaluation as mention in the previous sentence.