README.md

---
license: other
tags:
- ai-generated
- text-to-image
- prompt-engineering
- chinese
- image-caption
---

# Telegram AI Image Dataset — Cleaned for VLM LoRA Training

A cleaned dataset of **1,050 AI-generated images** with their generation prompts, collected from a Chinese Telegram channel focused on GPT-Image-2 prompt engineering.

Each image is paired with a structured generation prompt in Chinese/English. Prompts have been cleaned of channel boilerplate — no bot instructions, hashtags, source credits, model name prefixes, emoji title lines, or channel footer ads.

## Dataset Structure

| Column | Type | Description |
|--------|------|-------------|
| `message_id` | int64 | Telegram message ID |
| `datetime` | string | Message timestamp |
| `width` | int64 | Image width in pixels |
| `height` | int64 | Image height in pixels |
| `image` | binary | Embedded WebP image bytes |
| `text` | string | Cleaned generation prompt (Chinese/English) |

## Cleaning Applied

- **Removed 18 noise columns** — Telegram metadata, author info, file paths, entity data, etc.
- **Batch grouping** — Consecutive same-generation images grouped by matching dimensions; prompts propagated to all images in the batch.
- **Noise removal** — Stripped model name prefixes (`GPT-Image-2｜`), emoji title lines, bot instructions (`直接在 Bot 里输入提示词`), hashtags, source credits, channel footer ads (`VPN推荐`, `教程目录`, `邪修频道`).
- **Ad filtering** — Removed promotional posts, channel announcements, and non-prompt content.
- **Remaining** — 1,050 rows from 321 prompt groups, typically 2–4 images per prompt.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("GCStream/telegram-channel-dataset", split="train")
print(ds[0]["text"])  # Clean prompt
print(ds[0]["image"]) # PIL image
```

## Notes

- Images are embedded as binary WebP in the parquet `image` column.
- Prompts are primarily Chinese with English keywords, structured as field-value formats (任务, 主体, 场景, 光线, 镜头, 风格, etc.).
- Suitable for VLM fine-tuning (e.g., FLUX, SD3, DeepFloyd), prompt engineering analysis, and image-caption training.