Magnifico reposted this
Can we pre-train LLMs with Retrieval Augmentation? 🤔 RETRO was a research by Google DeepMind, which included retrieval into the pre-trainng process. Now NVIDIA continues this research by scaling RETRO to 48B, where they continued pretraining a 43B GPT model on an additional 100 billion tokens using the Retrieval augmentation method by retrieving from 1.2 trillion tokens. 🤯 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻: 1️⃣ Pretrain an LLM using the next token prediction on 1.1T token 2️⃣ Continue pretraining of the LLM with retrieval augmentation (Retro-fitting) on an additional 100 billion tokens while retrieving from whole pretraining dataset. 3️⃣ Instruction tune the model, with retrieval augmentation. Only update the weights of the decoder. 𝗜𝗻𝘀𝗶𝗴𝗵𝘁𝘀: 📈 Continued pretraining of LLMs with retrieval yields better decoders for QA. 🚀 Up to 10% improvements doing retrieval augmented pretraining 📚 Pretraining with retrieval improves the incorporating of context 🧮 Retrieval database had 19 billion chunks, with each chunk containing 64 tokens 🐌 Retrieval pretraining is slower and more complex due index and retrieval step Check out the full paper: https://lnkd.in/d2R4zsQ5 Remember that these are just my personal findings. Make sure always to conduct your own research and analysis. 🤗
This is mind blowing
Only 64 token per chunk? But it is interesting to find that a model can be trained for the whole process of RAG.
Very cool! What was the cost on this?
Thank you for the post - I am thinking Domain Specific Retrieval Augmentation :)
Retrieval Augmented Pretraining RAP is cool but have you heard of RASTA? Rastafari!
Huge scalable "fine tune", too expensive.$_$
This is really insightful. Thanks for sharing
🤯🤯🤯
Developing Solutions for M+ Figure Organizations | Entrepreneur | Founder | AI/ML
10moInsightful. I'm also doing the practices of fine-tuning of model (llama) with retrieval but using Lora instead of decoder-only. The objective is to make the model to be able to work with real-time data and provide the answers on fresh data while having dialogue behaviour.