---
dataset_info:
  features:
  - name: Source
    dtype: string
  - name: Date
    dtype: int64
  - name: Text
    dtype: string
  - name: Token_count
    dtype: int64
  splits:
  - name: train
    num_bytes: 8122744210
    num_examples: 6366648
  download_size: 3707767805
  dataset_size: 8122744210
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
pretty_name: Project_CodeNet
size_categories:
- 1M<n<10M
task_categories:
- text-generation
language:
- code
license: other
---

# Project_CodeNet

## Overview

This dataset is constructed from the **Project CodeNet** corpus, consisting of competitive programming submissions collected from online judges.

We extract a large-scale code corpus designed for pretraining language models, with a focus on:
- clean executable code
- temporal metadata (submission time)
- minimal preprocessing to preserve the original distribution

---

## Dataset Statistics

- **Total samples:** ~6.37M
- **Total tokens:** ~3.06B
- **Average tokens per sample:** 480.44

### Token Length Distribution
- P50: 162 tokens
- P90: 679 tokens
- P95: 1035 tokens
- P99: 2702 tokens
---

## Construction

### Source
- Project CodeNet https://github.com/IBM/Project_CodeNet

### Filtering Rules

We apply the following steps:

1. **Keep only Accepted submissions**
   - Removes incorrect or incomplete code.

2. **Deduplication at metadata level**
   - For each `(problem_id, user_id, language)`, keep the **last accepted submission**
   - This approximates the user's final solution

3. **No content-based deduplication**
   - Similar solutions across users are preserved
   - Reflects real-world submission distribution

4. **No balancing**
   - Language and temporal distributions are kept as-is

---

## Fields

Each sample contains:

| Field | Description |
|------|------------|
| `Source` | Dataset name (`Project_CodeNet`) |
| `Date` | Submission year |
| `Text` | Source code |
| `Token_count` | Token count computed using `tiktoken` |

---

## Tokenization

- Tokenizer: `tiktoken`
- Encoding: `cl100k_base`

---

## Distribution Characteristics

### Language Distribution
The dataset is highly skewed toward C++:

- C++ dominates (~60%)
- Python is the second largest (~23%)
- Other languages form a long tail

### Temporal Distribution
The dataset is heavily concentrated in recent years:

- Majority of samples from **2019–2020**
- Reflects real submission activity in CodeNet

---

## Important Notes

- This dataset preserves the **original submission distribution** of CodeNet.
- It is **not balanced** across languages or time.
- It is primarily composed of **competitive programming code**, which may differ from production software code.
- Some level of **near-duplicate solutions** exists due to similar problem-solving strategies.

---

## Intended Use

- Pretraining code language models
- Studying temporal evolution of programming patterns
- Benchmarking under real-world distribution settings

---

## Limitations

- Not representative of general software engineering code
- Strong bias toward:
  - competitive programming tasks
  - algorithmic problem solving
- Language and temporal imbalance

---

## License

Please refer to the original **Project CodeNet** dataset for licensing details.

---

## Citation

If you use this dataset, please cite Project CodeNet:

@article{puri2021project,
title={Project CodeNet: A Large-Scale AI for Code Dataset for Learning a Diversity of Coding Tasks},
author={Puri, Ruchir and others},
year={2021}
}