News
EveryLLM

English first

Trained mostly on English. 206 models on EveryLLM so far.

English-first models were trained mostly on English text, because English dominates the web, books, code comments and the research papers that set the agenda. Most early language models, from BERT and GPT-2 to Chinchilla and the first Llama, fall in this group, and many small and research models still do. They often handle other languages, but with clearly lower quality.

Where the data came from

BERT (2018) learned from English Wikipedia and a large set of unpublished books. GPT-2 (2019) used WebText, about 40 GB of text from 8 million web pages linked from Reddit. EleutherAI's Pile (2020), an 825 GiB collection weighted towards English academic papers, books and code, trained GPT-J and Pythia. Common Crawl, a free archive of the web gathered since 2008, supplies the bulk of most training sets.

Why it still matters

English-first training shapes what models know and how they speak: they may default to American spelling, law and culture, and do worse on questions about other places. Many benchmarks exist only in English. Labs now add far more non-English data, but English remains the largest share for most general models, and research models built to study training, such as OLMo, often stay English to keep the experiment clean.

Famous english first models

Still to be written: OLMo (language model).

Further reading

fnSparrow 🇬🇧 United KingdomSparrow DeepMind's rule-following chatbot that cites web evidence for its answersAssistant · Research only · 2022 Chinchilla 🇬🇧 United KingdomChinchilla A 70-billion-parameter DeepMind model built to test how models should scale.Research · Research only · 2022 Stanford Alpaca 🇺🇸 United StatesStanford Alpaca A cheap instruction-tuned LLaMA 7B built at Stanford in 2023Research · Research only · 2023 fnGato 🇬🇧 United KingdomGato A single transformer network trained to do many tasks, from chat to robot controlMultimodal · Research only · 2022 Gopher 🇬🇧 United KingdomGopher A 280 billion parameter Transformer language model from DeepMind, studied for how scale changes ability.Research · Research only · 2021 DialoGPT 🇺🇸 United StatesDialoGPT A GPT-style model trained on Reddit threads to write chat repliesGeneral · Open weights · 2019 OPT 🇺🇸 United StatesOPT Meta AI's family of open decoder-only language models from 125M to 175B parametersGeneral · Conditional licence · 2022 AlphaCode 🇬🇧 United KingdomAlphaCode A code generation system that solves competitive programming problems.Code · Research only · 2022 Orca 2 🇺🇸 United StatesOrca 2 A small LLaMA-2-based model trained on synthetic data to reason.Small · Conditional licence · 2023 fnToolformer 🇺🇸 United StatesToolformer A language model that learns on its own when to call external tools through APIs.Research · Research only · 2023 fnWebGPT 🇺🇸 United StatesWebGPT A GPT-3 model fine-tuned to answer questions while browsing the webGeneral · Research only · 2021 Med-PaLM 🇺🇸 United StatesMed-PaLM A medical-domain language model aligned from Flan-PaLM using instruction prompt tuning.Domain · Research only · 2022 Minerva 🇺🇸 United StatesMinerva A language model trained further on technical content to solve maths and science problemsResearch · Research only · 2022

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.