---
dataset_info:
  features:
  - name: asm
    sequence:
      sequence: string
  - name: desc
    sequence: string
  splits:
  - name: train
    num_bytes: 34436
    num_examples: 257
  download_size: 24042
  dataset_size: 34436
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---


# Evaluation Dataset for Assembly Code Search

This is the evaluation dataset used in [Virtual Compiler Is All You Need For Assembly Code Search]().

## Source

Various ways used to construct it:

- For Windows and macOS, we manually reverse-engineered many closed-source applications and constructed natural language queries for the assembly code.
- For Linux, we extract the source code and corresponding docstrings. We then verify the relevance between the source code, docstrings, and the corresponding assembly code. Then we filtered out the pairs with low relevance.

## Usage

This dataset is metadata for indexing only, as the binary is too large to put into a dataset.

For each row in the dataset, there are two columns:

- The `asm` column contains one or multiple tuples (a list of length equals two), the first item is the binary name and the second item is the address of the assembly code in that binary.
- The `desc` column contains one or multiple strings, each strings represents a natural language query.

The binaries can be downloaded at [here](https://cloud.vul337.team:8443/s/ygNnr4zm6KWL2jN) or [here (hf)](https://huggingface.co/datasets/elsagranger/AssemblyCodeSearchEval/tree/main/eval-binaries)