|
|
|
|
|
--- |
|
|
license: cc-by-sa-4.0 |
|
|
tags: |
|
|
- competitive-programming |
|
|
- code-ranking |
|
|
- llm-benchmark |
|
|
- code-efficiency |
|
|
- aizu-online-judge |
|
|
--- |
|
|
|
|
|
# AOJ-CodeRank-Benchmark: Hybrid Efficiency Ranking Benchmark Dataset |
|
|
|
|
|
## 1. Overview |
|
|
|
|
|
This dataset (AOJ-CodeRank-Benchmark) was created to evaluate the capability of **Large Language Models (LLMs)** in **code efficiency ranking tasks** using a high-quality, structured benchmark. |
|
|
|
|
|
The dataset is built entirely on code submission records from **Aizu Online Judge (AOJ)**, strictly adhering to the principle of **correctness first, efficiency second**. |
|
|
|
|
|
* **Problem Scope**: ALDS1 (Fundamental Algorithms), DSL/GRL/CGL (Advanced Data Structures/Graphs), and Volume 0000-3299 (Classic Contest Problems). |
|
|
* **Core Feature**: **Eliminates** 0ms submissions and low-quality/non-unique submissions, ensuring true time differentiation across all data groups. |
|
|
|
|
|
## 2. Data Structure |
|
|
|
|
|
The dataset uses the **JSON Lines (.jsonl)** format. Each line represents a single Task Group object. |
|
|
|
|
|
**Structure Preview (Candidates):** |
|
|
|
|
|
| Field Name | Type | Description | |
|
|
| :--- | :--- | :--- | |
|
|
| `submission_id` | string | Unique Submission ID. | |
|
|
| `code_snippet` | string | The complete C++ source code. | |
|
|
| **`accuracy`** | float | **Accuracy Score** (0.0 to 1.0). | |
|
|
| `time_ms` | integer | Actual Execution Time (in milliseconds). | |
|
|
| **`score_of_the_acc`** | float | **Normalized Efficiency Score** (Range -2.0 to 0.0). | |
|
|
| **`final_rank`** | integer | **Final Competition Rank** (1, 2, 3...). | |
|
|
|
|
|
## 3. Ground Truth (GT) Scoring and Ranking Logic 🏆 |
|
|
|
|
|
The LLM's objective is to predict the `final_rank`. This ranking is derived from a unique two-tiered system: |
|
|
|
|
|
### Phase I: Efficiency Score (`score_of_the_acc`) |
|
|
|
|
|
This score is a purely performance-based metric, calculating the normalized inverse sum of Time and Memory costs within the task group. |
|
|
|
|
|
$$ ext{Score} = -( ext{Norm\_Time} + ext{Norm\_Memory})$$ |
|
|
|
|
|
*(Note: Score is between -2.0 and 0.0. A score closer to 0.0 is better.)* |
|
|
|
|
|
### Phase II: Final Ranking (`final_rank`) Mechanism |
|
|
|
|
|
The final rank is determined by a lexicographical sort (Standard Competition Ranking) using the following priority: |
|
|
|
|
|
1. **Primary Sort Key (Accuracy)**: **`accuracy`** (Descending). |
|
|
2. **Secondary Sort Key (Efficiency)**: **`score_of_the_acc`** (Descending). |
|
|
|
|
|
**Tie-Breaking**: Submissions with identical Accuracy and Efficiency Score receive the same rank (1-2-2-4 rule). |
|
|
|
|
|
--- |
|
|
|
|
|
### 4. Usage Example |
|
|
|
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
# Load the dataset and access the candidates list |
|
|
dataset = load_dataset("Slime/AOJ-CodeRank-Benchmark", data_files="train.jsonl", split="train") |
|
|
|
|
|
# The LLM sorting algorithm will receive task['candidates'] for ranking |
|
|
for task in dataset: |
|
|
candidates = task['candidates'] |
|
|
# Algorithm generates predicted_rank for candidates |
|
|
# Evaluation compares predicted_rank against ground_truth['final_rank'] |
|
|
``` |
|
|
|
|
|
### 5. Acknowledgments |
|
|
Original submission records and problem context are sourced from Aizu Online Judge (AOJ). |
