sdeakin/LLM-Tagged-GoEmotions

---
license: cc-by-4.0
task_categories:
- text-classification
- sentence-similarity
- feature-extraction
language:
- en
tags:
- emotion-classification
- text-classification
- explanations
- rationales
- goemotions
- GoEmotions
- synthetic
- llm-generated
- natural-language-processing
- emotions
- affect
pretty_name: 'LLM-Tagged GoEmotions: Llama 3 Labeling of GoEmotions'
size_categories:
- 100K<n<1M
---

# Dataset Card for **LLM-Tagged-GoEmotions**

## Dataset Summary

`LLM-Simple-Emotions.jsonl` contains **211,225 synthetic emotion annotations** generated from the original GoEmotions corpus.

Each Reddit utterance is re-annotated using **`llama3:instruct`** (via Ollama) with the **Simple Level-1 Prompt**, which instructs the model to:

* Predict the **primary emotion label(s)** (from GoEmotions)
* Provide a **natural-language explanation** of *why* those emotions were tagged

This dataset is ideal for:

* Single-label and multi-label emotion classification
* Training models that use **rationale/explanation supervision**
* Studying LLM emotional reasoning over text

---

## Supported Tasks

### **Emotion Classification**

Use:

* `data.labels`

### **Explanation Modeling (Optional)**

Use:

* `data.explanation`
To train models to generate text rationales or explanations.

---

## Languages

* **English (`en`)**

---

## Dataset Structure

### **Example Record**

```json
{
"src_id": "l1_0",
"model": "llama3:instruct",
"provider": "ollama-local",
"prompt": "simple_level1",
"text": "That game hurt.",
"data": {
"labels": ["disappointment"],
"explanation": "The speaker expresses regret and sadness about the outcome of the game, indicating disappointment."
}
}
```

---

## Size & Splits

* **Total entries:** 211,225
* **Splits:** Single combined dataset (`train` only)
Users may create custom train/validation/test splits.

---

## Data Collection & Processing

### **Source**

* Original GoEmotions dataset (`CC BY 4.0`)

### **Generation Pipeline**

1. Load each GoEmotions utterance.
2. Apply the Simple Level-1 prompt to `llama3:instruct`.
3. Extract:

* Emotion label(s)
* Explanation text
4. Save the structured result into JSONL.

### **Post-Processing**

Minimal cleanup:

* Remove malformed outputs and nonsensical labels
* Normalize labels
* Ensure text + explanation are present

---

## Known Limitations

### **Model Bias**

* Labels and explanations depend on Llama-3’s internal reasoning and biases.
* Explanations may be overly confident or simplistic.

---

## Usage

### **Direct JSONL Reading**

```python
import json

with open("LLM-Simple-Emotions.jsonl", "r", encoding="utf-8") as f:
for line in f:
record = json.loads(line)
print(record["text"], record["data"]["labels"], record["data"]["explanation"])
```

### **Load with Hugging Face Datasets**

```python
from datasets import load_dataset

ds = load_dataset(
"json",
data_files="LLM-Simple-Emotions.jsonl",
split="train"
)
```

---

## Citation

Please cite both the **original GoEmotions dataset** and this LLM-generated extension:

```bibtex
@article{demszky2020goemotions,
title={GoEmotions: A Dataset of Fine-Grained Emotions},
author={Demszky, Dorottya and et al.},
journal={ACL},
year={2020}
}

@dataset{LLM-Tagged-GoEmotions,
title={LLM-Tagged GoEmotions: Llama 3 Labeling of GoEmotions},
author={Sheryl D. and contributors},
year={2025},
url={https://huggingface.co/datasets/sdeakin/LLM-Tagged-GoEmotions}
}
```

---

## Contact

For questions or issues, please open an issue on the dataset repository or contact me.

license: 知识共享署名4.0（CC BY 4.0）
task_categories:
- 文本分类（text-classification）
- 句子相似度（sentence-similarity）
- 特征提取（feature-extraction）
language:
- 英语（en）
tags:
- 情感分类（emotion-classification）
- 文本分类（text-classification）
- 解释（explanations）
- 理由依据（rationales）
- GoEmotions
- GoEmotions
- 合成数据（synthetic）
- LLM生成数据（llm-generated）
- 自然语言处理（natural-language-processing）
- 情感（emotions）
- 情感状态（affect）
pretty_name: "LLM标注GoEmotions数据集：Llama 3对GoEmotions的标注结果"
size_categories:
- 100K<n<1M
---

# LLM标注版GoEmotions数据集卡片

## 数据集概览

`LLM-Simple-Emotions.jsonl` 包含源自原始GoEmotions语料库生成的**211,225条合成情感标注**。每条Reddit平台发言均通过Ollama调用`llama3:instruct`模型，配合**简单一级提示词**进行重新标注，该提示词要求模型完成两项任务：
* 预测来自GoEmotions的**主要情感标签**（支持单标签或多标签）
* 以自然语言形式解释**标注该情感的依据**

本数据集适用于以下场景：
* 单标签与多标签情感分类任务
* 训练具备**理由/解释监督能力**的模型
* 研究大语言模型（Large Language Model，LLM）对文本的情感推理过程

---

## 支持任务

### 情感分类任务
使用字段：
* `data.labels`

### 解释建模（可选）
使用字段：
* `data.explanation`
用于训练能够生成文本理由或解释的模型。

---

## 语言

* **英语（`en`）**

---

## 数据集结构

### 示例数据记录

json
{
"src_id": "l1_0",
"model": "llama3:instruct",
"provider": "ollama本地部署版",
"prompt": "simple_level1",
"text": "That game hurt.",
"data": {
"labels": ["disappointment"],
"explanation": "The speaker expresses regret and sadness about the outcome of the game, indicating disappointment."
}
}

---

## 规模与数据集拆分

* **总条目数：** 211,225
* **数据集拆分：** 仅包含单一合并训练集（`train`），用户可自行划分自定义的训练/验证/测试集。

---

## 数据收集与处理流程

### 数据来源

* 原始GoEmotions数据集（知识共享署名4.0许可协议）

### 生成流程

1. 加载每条原始GoEmotions发言
2. 对`llama3:instruct`模型应用简单一级提示词
3. 提取以下内容：
* 情感标签
* 解释文本
4. 将结构化结果保存为JSONL格式文件。

### 后处理步骤

仅进行极简清理操作：
* 移除格式错误的输出与无意义标签
* 标准化标签格式
* 确保文本与解释字段均完整存在

---

## 已知局限性

### 模型偏见

* 生成的标签与解释结果依赖Llama-3的内部推理逻辑与固有偏见
* 生成的解释可能过于绝对或过于简化

---

## 使用方法

### 直接读取JSONL文件

python
import json

with open("LLM-Simple-Emotions.jsonl", "r", encoding="utf-8") as f:
for line in f:
record = json.loads(line)
print(record["text"], record["data"]["labels"], record["data"]["explanation"])

### 使用Hugging Face Datasets库加载

python
from datasets import load_dataset

ds = load_dataset(
"json",
data_files="LLM-Simple-Emotions.jsonl",
split="train"
)

---

## 引用信息

请同时引用原始GoEmotions数据集与本次LLM生成的扩展数据集：

bibtex
@article{demszky2020goemotions,
title={GoEmotions: A Dataset of Fine-Grained Emotions},
author={Demszky, Dorottya and et al.},
journal={ACL},
year={2020}
}

@dataset{LLM-Tagged-GoEmotions,
title={LLM-Tagged GoEmotions: Llama 3 Labeling of GoEmotions},
author={Sheryl D. and contributors},
year={2025},
url={https://huggingface.co/datasets/sdeakin/LLM-Tagged-GoEmotions}
}

---

## 联系方式

如有疑问或问题，请在数据集仓库中提交Issue或联系作者。