Open LLMs that blend instructions with crawled ChatGPT chats

What it does FreedomIntelligence’s LLM Zoo is a research project that trains and distributes open large language models—chiefly Phoenix and Chimera—alongside the data and benchmarks used to build them. It packages CLI inference tools, training code, and evaluation tables comparing its own models against contemporaries like Alpaca, Vicuna, and ChatGPT. The stated goal is to make ChatGPT-like assistants accessible across languages and countries rather than locked inside a single company.

The interesting bit The project’s core technical bet is that language models need both instruction tuning and conversation data to feel useful—Alpaca supplied the former, Vicuna the latter, and LLM Zoo claims to be one of the earlier efforts to bake both into the same recipe. It also wears its politics on its sleeve, framing the work as an explicit attempt to break “AI supremacy” and prevent any single entity from monopolizing the direction of artificial intelligence.

Key highlights

• Releases two main model families: Phoenix (7B, BLOOMZ-based, multilingual) and Chimera (7B/13B, LLaMA-based, focused on Latin and Cyrillic languages)
• Publishes the training data, including multilingual instruction sets and crawled multi-round ChatGPT conversations
• Provides an evaluation benchmark that scores open models against ChatGPT and other contemporaries in a single comparison table
• Ships training code and a CLI inference tool that downloads Phoenix weights automatically from Hugging Face
• Also spawned CAMEL (HuatuoGPT), a Chinese medical model hosted in a separate repository

Caveats

• Chimera is distributed as delta weights only, so you must source the original LLaMA weights yourself and run a merge script before using it
• Several promised models—Legal GPT, vision-language variants, and retrieval-augmented systems—are listed as “coming soon” with no release date given
• The README’s evaluation benchmark table is mostly a static snapshot from April 2023, so it is unclear how actively the leaderboard is maintained

Verdict Worth a look if you are researching multilingual instruction tuning or need a reproducible pipeline for training 7B-scale chat models. Skip it if you are hunting for a drop-in replacement for GPT-4; the project is explicitly early-stage and academic in scope.