metadata

license: mit
task_categories:
  - text-generation
  - text2text-generation
language:
  - en
pretty_name: BluePrint
size_categories:
  - 1M<n<10M
configs:
  - config_name: 2_clusters
    data_files:
      - split: full
        path: processed_2_clusters/cluster_*.jsonl
    default: true
  - config_name: 25_clusters
    data_files:
      - split: full
        path: processed_25_clusters/cluster_*.jsonl
  - config_name: 100_clusters
    data_files:
      - split: full
        path: processed_100_clusters/cluster_*.jsonl
  - config_name: 1000_clusters
    data_files:
      - split: full
        path: processed_1000_clusters/cluster_*.jsonl

📘 BluePrint

BluePrint is a large-scale dataset of social media conversation threads designed for evaluating and training LLM-based social media agents. It provides realistic, thread-structured data clustered into representative user personas at various levels of granularity.

✅ Key Features

Thread-Based Structure: Each example is a list of messages representing a user thread.

Persona Clustering: Users are clustered into 2, 25, 100, and 1000 representative personas to enable safe and scalable simulations.

Behavioral Signals: Each message includes relative timestamps, user ID, and social actions such as likes, reposts, and replies.

🎯 Use Cases

Text Generation

Next-Action Prediction

Behavioral Simulation

Persona Modeling

📝 Dataset Structure

Each record contains:

relative_integer_time: Relative timestamp of the message.
text: Text content of the message.
actions: Next action (bool). Will have the following:
- like, unlike, repost, unrepost, follow, unfollow, block, unblock, post_update, post_delete, quote, post, reply
user_id: Anonymized user identifier.

Example:

{
    "thread": [
        {
            "relative_integer_time": 9884584,
            "text": "BTW, Mark Cuban is the second-most followed person on Bluesky. Back in 2017 there was a lot of talk about who could be the Democrats' Trump, and plenty of chest-puffing in response that there couldn't be one. Cuban is it. He's not as dim and he's not a sociopath but he is talking out of his ass.",
            "actions": {
                "like": false,
                "unlike": false,
                "repost": false,
                "unrepost": false,
                "follow": false,
                "unfollow": false,
                "block": false,
                "unblock": false,
                "post_update": false,
                "post_delete": false,
                "quote": false,
                "post": true,
                "reply": false
            },
            "user_id": "7ebac0bdfe8f0925bf1e5dd672e3288e3b1e05ba0620285f7d7f7ed89e33ee9e"
        }
    ]
}

Statistics of each cluster split can be found at the folder stats/.

🗂️ Configurations

Config Name	Description	Size (examples)
`2_clusters`	2 representative user personas	6.8M
`25_clusters`	25 representative user personas	6.8M
`100_clusters`	100 representative user personas	6.8M
`1000_clusters`	1000 representative user personas	6.8M

💾 Loading Example

from datasets import load_dataset

dataset = load_dataset("ComplexDataLab/BluePrint", name="2_clusters")

for i, example in enumerate(dataset["full"]):
    print(f"Example {i}: {example}")
    break

Example 0: {'thread': [{'relative_integer_time': 9884584, 'text': "BTW, Mark Cuban is the second-most followed person on Bluesky. Back in 2017 there was a lot of talk about who could be the Democrats' Trump, and plenty of chest-puffing in response that there couldn't be one. Cuban is it. He's not as dim and he's not a sociopath but he is talking out of his ass.", 'actions': {'like': False, 'unlike': False, 'repost': False, 'unrepost': False, 'follow': False, 'unfollow': False, 'block': False, 'unblock': False, 'post_update': False, 'post_delete': False, 'quote': False, 'post': True, 'reply': False}, 'user_id': '7ebac0bdfe8f0925bf1e5dd672e3288e3b1e05ba0620285f7d7f7ed89e33ee9e'}]}

📄 License

This dataset is released under the MIT License.