Curating the chaos of chatbot training data

What it does

This repository is a curated index of publicly available instruction-tuning, chat, and RLHF datasets, sorted by size and tagged with language and license. It also includes a lightweight Python script to merge your chosen subsets into a single dataset and push the result to the Hugging Face Hub.

The interesting bit

The value is in the aggregation: instead of scraping Hugging Face individually, you get a single table spanning everything from 1K math examples to multi-million-entry conversation dumps. The included preprocess.py turns that curation into a custom training mix without writing your own merge logic.

Key highlights

• Covers the full spectrum of modern LLM training needs: math, coding, safety, multilingual dialogue, multimodal prompts, and human preference pairs.
• Datasets range from tiny specialized collections (1K examples) to massive blends (3M+ entries), with size and licensing noted for quick triage.
• Includes a simple mixing utility to combine selected datasets and upload them directly to Hugging Face.
• Tracks provenance and license terms—handy for spotting non-commercial or OpenAI-terms restrictions at a glance.

Caveats

• Many entries list “-” for license, leaving legal clarity as an exercise for the reader.
• The mixing script is barely documented; the README shows a one-line invocation with no explanation of deduplication, formatting, or validation.
• No quality benchmarks or contamination checks are provided, so you are trusting the upstream sources blindly.

Verdict

Worth bookmarking if you are building a custom chat model and need a starting point for data sourcing. Skip it if you expect ready-to-train, cleaned, and benchmarked dataset recipes.