The raw material index for DIY chatbots

What it does

This repository is a curated markdown catalog of open datasets used to fine-tune instruction-following LLMs—think Alpaca, Dolly, OpenAssistant, and multimodal variants like LLaVA. Each entry records the dataset size, language, task type, generation method (human-written, self-instruct, mixed, or collection), and license. It is essentially a lookup table for researchers trying to avoid training on data that is either toxic or legally toxic.

The interesting bit

The value is in the taxonomy. The README tags every dataset by provenance—distinguishing human-generated data from text-davinci-003 or GPT-4 distillations—and flags licenses ranging from permissive (Apache, MIT) to strictly non-commercial. That distinction matters when your “open” model is headed for production, and a dedicated section catalogs RLHF preference datasets and visual-instruction pairs that are harder to stumble across.

Key highlights

• Covers text, multimodal, and RLHF datasets (e.g., Anthropic hh-rlhf, LLaVA, ShareGPT52K) with rough size estimates.
• Tags languages (EN, CN, ML), generation method (HG, SI, MIX, COL), and software license for each entry.
• Notes the synthetic data pipeline used where available—GPT-4, text-davinci-003, or human curation.
• Includes a contribution template and links to a companion list of training codebases.

Caveats

• The list is a static markdown file, not a searchable database, so discovery still requires scrolling.
• Formatting is inconsistent and the README appears truncated in places (e.g., the GPTeacher entry cuts off after the “paper” field).
• Some entries contain copy-paste errors or conflicting metadata, so treat the numbers as approximate.

Verdict

Worth bookmarking if you are fine-tuning a chat model and need to compare dataset provenance and licensing at a glance. Skip it if you want training scripts or programmatic data loaders.