Also, it's worth noting that RAG (retrieval augmented generation) doesn't fix the problem. See those nice links into NYC web pages? Not stopping the system from *making shit up*. (Second column is chatbot response, third is journalist's report on the actual facts.) >>
LLMS
-
LLMs Fundamentally Designed to Generate Inaccurate Content
By
–
It seems to bear repeating: chatbots based on large language models are designed to *make shit up*. This isn't a fixable bug. It's a fundamental mismatch between tech and task. >>
-
Strategic Data Use in Alignment Tuning: sDPO Approach
By
–
Nice introduction of sDPO: Strategic Data Use in Alignment Tuning. https://
linkedin.com/pulse/e1i8-win
dows-worship-ravi-naukarkar-aed5f/?trackingId=ztABINS8QWUZWFhEOAqLeg
… "Enhanced Alignment: Leveraging preference data in stages, sDPO ensures models align more closely with preferences." "Superior Performance: The stepwise approach doesn’t just fine-tune; -
Building ToolLLM With LangGraph.js: API Integration
By
–
New YouTube Video: Building ToolLLM With LangGraph.js Inspired by the ToolLLM paper by @TsingYoga
, which introduces an innovative approach to utilizing a vast array of APIs to address user queries. The key lies in their hierarchical API retriever, which quickly -
Dataverse: High-Quality Data Filtering for LLM Training
By
–
Looking for high-quality data for your LLM training? Try https://
github.com/UpstageAI/data
verse
…. It filters out low-quality data and generates quality text from large inputs. #solarllm is powered by Dataverse! -
Pretraining Vision and LLMs in Python on AWS
By
–
Pretrain Vision and #LLMs (Large Language Models) in #Python — Techniques for building & deploying foundation models on #AWS: http://
amzn.to/3pT9seD via @PacktPublishing ————
#AI #MachineLearning #DeepLearning #MLOps #BigData #DataScience #NLProc #ComputerVision #DataScientists -
Key Challenges in Building RAG Solutions for Business
By
–
Building RAG solutions is becoming more common and important in the business world. But there are certain complexities and challenging problems such as:
•How to differentiate between similar-sounding documents
•Rank disparate data sources
•Re-write query to ensure you are -
Next Word Prediction: How Language Models Learn Knowledge
By
–
Predicting the next word "only" is sufficient for language models to learn a large body of knowledge that enables then to code, answer questions, understand many topics, chat, and so on.
— Nando de Freitas (@NandoDF) 29 mars 2024
This is clear to many researchers now, and there are nice tutorials on why this works by… pic.twitter.com/L95G6hWRNVPredicting the next word "only" is sufficient for language models to learn a large body of knowledge that enables then to code, answer questions, understand many topics, chat, and so on. This is clear to many researchers now, and there are nice tutorials on why this works by
-
Fine-tuning Jamba Chat with PEFT adapter merging
By
–
Trained a Jamba-Chat w/ PEFT. Having a lot of trouble merging the weights for some reason. If anyone can help out, we'll have a chat version of Jamba! Here are the adapters:
-
New 9,160 Test Cases Dataset for Prompt Injection Evaluation
By
–
The most interesting thing about this paper is the eval they released with it, containing 9,160 test cases for prompt injection https://
github.com/egozverev/Shol
d-It-Be-Executed-Or-Processed
… Explore it in Datasette Lite here: https://
lite.datasette.io/?json=https://
github.com/egozverev/Shold-It-Be-Executed-Or-Processed/blob/main/SEP_dataset/SEP_dataset.json#/data/SEP_dataset
…