Transformers and attention

13 milestones used this technique.

Mixtral 8x7B: sparse mixture-of-experts model released under Apache 2.0

Mixtral 8x7B, a sparse mixture-of-experts language model, was published in January 2024. It outperforms or matches Llama 2 70B and GPT-3.5 on most evaluated benchmarks while using only 13B active parameters per token, and both base and instruction-tuned versions were released under the Apache 2.0 licence.

FlashAttention-2 roughly doubles attention speed on A100 GPUs

Tri Dao's FlashAttention-2 improved GPU work partitioning to deliver roughly 2x the speed of FlashAttention, reaching 50–73% of the theoretical maximum FLOPs/s in the configurations benchmarked in the paper, and up to 225 TFLOPs/s per A100 in the specific GPT-style training configurations reported.

QLoRA enables finetuning of 65B-parameter models on a single 48GB GPU

Tim Dettmers and colleagues showed that a 65-billion-parameter language model could be finetuned on a single 48GB GPU without measurable quality loss on the benchmarks tested, by combining 4-bit quantisation with low-rank adapter training.

LLaMA matches leading models on public data with far fewer parameters

LLaMA, a collection of foundation language models from 7B to 65B parameters trained exclusively on publicly available data, was submitted to arXiv on 27 February 2023. The 13B model outperformed GPT-3 at 175B parameters on most benchmarks, and the weights were released to the research community.

BLOOM: open-access 176B-parameter multilingual language model

The BigScience Workshop released BLOOM, a 176-billion-parameter language model trained across 46 natural and 13 programming languages, made freely available under the Responsible AI License at a scale that had only recently begun to become accessible, and had not previously been available with multilingual training data.

Flan-PaLM: instruction finetuning scaled across tasks, model sizes and families

Researchers showed that finetuning large language models on instruction-phrased datasets improves performance across benchmarks, with Flan-PaLM 540B trained on 1.8K tasks scoring 75.2% on five-shot MMLU and outperforming its base model by 9.4% on average.

Flamingo: few-shot visual language model for interleaved images, video and text

Researchers introduced Flamingo, a family of Visual Language Models that could handle interleaved images, video and text, achieving state-of-the-art few-shot performance on many benchmarks without task-specific fine-tuning.

Chain-of-thought prompting improves step-by-step reasoning in large language models

Jason Wei and colleagues showed that prompting a large language model with worked examples of intermediate reasoning steps dramatically improved its performance on arithmetic, commonsense, and symbolic tasks, reaching state-of-the-art accuracy on a maths benchmark.

Instruction tuning lets a 137B model match or beat GPT-3 zero-shot

Jason Wei and colleagues showed that finetuning a 137B language model on over 60 NLP tasks described via natural language instruction templates produced a model, FLAN, that beat zero-shot GPT-3 on 20 of 25 tasks and surpassed few-shot GPT-3 on several benchmarks.

Microsoft Research released Turing Natural Language Generation (T-NLG), a 17-billion-parameter language model

In February 2020, Microsoft Research announced Turing Natural Language Generation (T-NLG), a 17-billion-parameter autoregressive language model trained using the Megatron-LM framework. At the time of release it was the largest publicly disclosed language model and achieved state-of-the-art results on question-answering and summarisation benchmarks.

BERT introduces masked bidirectional pre-training for language understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova introduced BERT, a language model that pre-trains on both left and right context simultaneously, achieving new best results on eleven natural language processing tasks.

OpenAI Releases GPT-1: Improving Language Understanding by Generative Pre-Training

In June 2018, Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever at OpenAI published 'Improving Language Understanding by Generative Pre-Training', introducing GPT-1, a 117-million-parameter Transformer pretrained on BooksCorpus via unsupervised language modelling and fine-tuned on downstream tasks, outperforming task-specific models on several NLP benchmarks.

Vaswani et al. introduce the Transformer, a sequence model built on self-attention alone

Ashish Vaswani and colleagues proposed the Transformer, a sequence model whose encoding and decoding relied on multi-head self-attention rather than recurrence or convolutions. It set new translation benchmarks while training faster than existing models.