---
dataset_info:
- config_name: edit
  features:
  - name: input
    dtype: string
  - name: target
    dtype: string
  - name: problem_id
    dtype: string
  splits:
  - name: train
    num_bytes: 56166875
    num_examples: 48386
  - name: val
    num_bytes: 3336062
    num_examples: 3338
  - name: test
    num_bytes: 857857
    num_examples: 794
  download_size: 365069
  dataset_size: 60360794
- config_name: generate
  features:
  - name: problem_id
    dtype: string
  - name: problem_description
    dtype: string
  splits:
  - name: train
    num_bytes: 1793963
    num_examples: 1262
  - name: val
    num_bytes: 96855
    num_examples: 69
  - name: test
    num_bytes: 60776
    num_examples: 49
  download_size: 37588
  dataset_size: 1951594
- config_name: generate_eval
  features:
  - name: problem_id
    dtype: string
  - name: runtimes
    sequence: float64
  - name: memories
    sequence: float64
  - name: num_sol
    dtype: int64
  splits:
  - name: test
    num_bytes: 770704
    num_examples: 48
  download_size: 147211
  dataset_size: 770704
configs:
- config_name: edit
  data_files:
  - split: train
    path: edit/train-*
  - split: val
    path: edit/val-*
  - split: test
    path: edit/test-*
- config_name: generate
  data_files:
  - split: train
    path: generate/train-*
  - split: val
    path: generate/val-*
  - split: test
    path: generate/test-*
- config_name: generate_eval
  data_files:
  - split: test
    path: generate_eval/test-*
---
# ECCO

Dataset from the paper "ECCO: Can We Improve Model-Generated Code Efficiency Without Sacrificing Functional Correctness?"

![teaser](https://github.com/user-attachments/assets/44659b06-3676-4deb-affb-2ec5f02787f6)

The dataset consists of 2 subsets `edit` and `generate` each with 3 splits (`train`, `val` and `test`).

Code repository: [https://github.com/CodeEff/ECCO](https://github.com/CodeEff/ECCO)

### Loading the dataset / benchmark 
```python
dataset = load_dataset('CodeEff/ECCO', 'edit') # For history-based editing setting
dataset = load_dataset('CodeEff/ECCO', 'generate') # For nl-instructed generation setting
```
These are used to generate code by each model across the 2 paradigms. We use the `test` split for the evaluation/results and the `train` and `val` splits for finetuning and few-shot prompting.

### Download the test cases 
```sh
mkdir data && cd data
wget https://huggingface.co/datasets/CodeEff/ECCO/resolve/main/test_cases.zip
unzip test_cases.zip
```

### Evaluation dataset
The dataset also consists of an additional 3rd subset `generate_eval` which consists of the runtime and memory of a spectrum of user solutions for each problem in the `test` split.     
This is used for the percentile evaluation of the **NL-instructed generation** paradigm.

### Data Sources 
Dataset is sourced from [IBM CodeNet](https://github.com/IBM/Project_CodeNet) which consists of primarily competetive programming solutions. 
This is further filtered for efficiency and correctness as described in our paper. 


### Citation 
```bib
@article{waghjale2024ecco,
  title={ECCO: Can We Improve Model-Generated Code Efficiency Without Sacrificing Functional Correctness?},
  author={Waghjale, Siddhant and Veerendranath, Vishruth and Wang, Zora Zhiruo and Fried, Daniel},
  journal={arXiv preprint arXiv:2407.14044},
  year={2024}
}
```