Train 1 trillion+ parameter models https://lightning.ai/docs/pytorch/stable/advanced/model_parallel.html https://docs.mosaicml.com/projects/composer/en/latest/trainer/using_the_trainer.html
Introducing MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs https://www.mosaicml.com/blog/mpt-7b
The Self-Supervised Learning Cookbook https://ai.facebook.com/blog/self-supervised-learning-practical-guide/
Best
https://www.youtube.com/watch?v=bZQun8Y4L2A
Pretraining
- LLaMA 65B (NC)
- MPT-7B
- Can do long context
- RedPajama 3B, 7B
- GeoV-9B
- RWKV-14B
DoReMi https://huggingface.co/papers/2305.10429
SFT
Reward Modeling
RLHF
Hard to work in production…
Easier to judge vs to write an example Less diverse outputs Base model is better for lists of N things
LIMA
https://arxiv.org/abs/2305.11206 https://medium.com/@WellAI/is-less-more-for-llms-a-critical-analysis-of-metas-lima-ff15de6b842e
SLiC-HF https://huggingface.co/papers/2305.10425
Principle-Driven Self-Alignment of Language Models from Scratch with Minimal Human Supervision https://huggingface.co/papers/2305.03047
AutoML-GPT https://arxiv.org/abs/2305.02499
System 2 Thinking https://github.com/kyegomez/tree-of-thoughts https://huggingface.co/papers/2305.10601
https://arxiv.org/abs/2210.03629 https://ai.googleblog.com/2022/11/react-synergizing-reasoning-and-acting.html
https://github.com/Significant-Gravitas/Auto-GPT
https://arxiv.org/abs/2211.01910
You are a world leading expert “Let’s work this out in a step by step way to be sure we have the right answer”
https://arxiv.org/abs/2302.04761 https://github.com/lucidrains/toolformer-pytorch https://vinija.ai/models/Toolformer/
Constrained Prompting
https://github.com/microsoft/guidance