IRC Disentangle

This dataset provides manually annotated IRC logs with response structures, allowing you to reconstruct tangled conversations. A valuable resource for training or evaluating models for understanding dialogue.

Download dataset

Size

77,563 JSON-annotated IRC messages with response graphs

Licence

CC-BY 4.0

Description

‍

The dataset IRC Disentangle offers over 77,000 messages from IRC channels, each annotated manually to indicate the response links between messages. It thus makes it possible to reconstruct mixed conversations typical of online chats. Each message comes with a raw version, an ASCII version, a tokenization, and response graphs showing the associated messages. The game is divided into channels and dates, with particular attention paid to the quality of the annotations.

‍

What is this dataset for?

‍

Train conversational deinterlacing models (disentanglement)
Test a chatbot's ability to identify threads in a disorganized flow
Improve tools for moderating and analysing online dialogues

‍

Can it be enriched or improved?

‍

Yes, the dataset can be enriched by adding other IRC channels, translating to other languages, or adding additional annotations (emotions, intentions, etc.). Graphs can also be used to generate realistic synthetic conversations, useful for fine-tuning dialogue models.

‍

🔎 In summary

Criterion	Evaluation
🧩 Ease of use	⭐⭐⭐✩✩ (Clear structure but requires familiarity with graphs)
🧼 Need for cleaning	⭐⭐⭐⭐⭐ (Low – data already normalized)
🏷️ Annotation richness	⭐⭐⭐⭐✩ (Very good – explicit connections between messages)
📜 Commercial license	✅ Yes (CC-BY 4.0)
👨‍💻 Beginner friendly	⚠️ Accessible but requires understanding of graphs
🔁 Fine-tuning ready	🎯 Yes, for dialogue disentanglement or multi-turn chatbots
🌍 Cultural diversity	⚠️ Medium – English only (Ubuntu logs)