Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
datasets:
|
| 4 |
+
- DaniilOr/SemEval-2026-Task13
|
| 5 |
+
metrics:
|
| 6 |
+
- accuracy
|
| 7 |
+
- f1
|
| 8 |
+
- precision
|
| 9 |
+
- recall
|
| 10 |
+
base_model:
|
| 11 |
+
- microsoft/unixcoder-base
|
| 12 |
+
library_name: transformers
|
| 13 |
+
tags:
|
| 14 |
+
- detection
|
| 15 |
+
- AI-generated
|
| 16 |
+
- transformers
|
| 17 |
+
- bert
|
| 18 |
+
---
|
| 19 |
+
## 🔍 Task Overview
|
| 20 |
+
|
| 21 |
+
The rise of generative models has made it increasingly difficult to distinguish machine-generated code from human-written code — especially across different programming languages, domains, and generation techniques.
|
| 22 |
+
|
| 23 |
+
**SemEval-2026 Task 13** challenges participants to build systems that can **detect machine-generated code** under diverse conditions by evaluating generalization to unseen languages, generator families, and code application scenarios.
|
| 24 |
+
|
| 25 |
+
The task consists of **three subtasks**:
|
| 26 |
+
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
+
### Subtask A: Binary Machine-Generated Code Detection
|
| 30 |
+
|
| 31 |
+
**Goal:**
|
| 32 |
+
Given a code snippet, predict whether it is:
|
| 33 |
+
|
| 34 |
+
- **(i)** Fully **human-written**, or
|
| 35 |
+
- **(ii)** Fully **machine-generated**
|
| 36 |
+
|
| 37 |
+
**Training Languages:** `C++`, `Python`, `Java`
|
| 38 |
+
**Training Domain:** `Algorithmic` (e.g., Leetcode-style problems)
|
| 39 |
+
|
| 40 |
+
**Evaluation Settings:**
|
| 41 |
+
|
| 42 |
+
| Setting | Language | Domain |
|
| 43 |
+
|--------------------------------------|-------------------------|------------------------|
|
| 44 |
+
| (i) Seen Languages & Seen Domains | C++, Python, Java | Algorithmic |
|
| 45 |
+
| (ii) Unseen Languages & Seen Domains | Go, PHP, C#, C, JS | Algorithmic |
|
| 46 |
+
| (iii) Seen Languages & Unseen Domains| C++, Python, Java | Research, Production |
|
| 47 |
+
| (iv) Unseen Languages & Domains | Go, PHP, C#, C, JS | Research, Production |
|
| 48 |
+
|
| 49 |
+
**Dataset Size**:
|
| 50 |
+
- Train - 500K samples (238K Human-Written | 262K Machine-Generated)
|
| 51 |
+
- Validation - 100K samples
|
| 52 |
+
|
| 53 |
+
**Data Format**
|
| 54 |
+
Each dataset contains the following fields:
|
| 55 |
+
- `code`: The code snippet
|
| 56 |
+
- `label`: The binary label (0 for human-written, 1 for machine-generated)
|
| 57 |
+
- `language`: The programming language of the snippet
|
| 58 |
+
|
| 59 |
+
Label mappings are provided in `task_A/label_to_id.json` and `task_A/id_to_label.json`.
|
| 60 |
+
|
| 61 |
+
**Evaluation Metric**
|
| 62 |
+
The primary evaluation metric for Subtask A is **Macro F1-score**. This metric ensures balanced performance across both classes.
|
| 63 |
+
|
| 64 |
+
**Submission Format**
|
| 65 |
+
Participants must submit a `.csv` file with the following columns:
|
| 66 |
+
- `id`: Unique identifier for each code snippet
|
| 67 |
+
- `label`: Predicted label (0 or 1)
|
| 68 |
+
|
| 69 |
+
A sample submission file is available in the `task_A/` folder.
|
| 70 |
+
|
| 71 |
+
**Baseline Models**
|
| 72 |
+
Baseline implementations for Subtask A are provided in the `baselines/` directory. These include starter code and pre-trained checkpoints for models such as GraphCodeBERT and UniXcoder.
|
| 73 |
+
|
| 74 |
+
**Restrictions**
|
| 75 |
+
- **No external training data**: Use only the provided datasets.
|
| 76 |
+
- **No specialized AI-generated code detectors**: General-purpose code models (e.g., CodeBERT, StarCoder) are allowed.
|