Anyscale’s Post

Anyscale LLM Router template #template #llmrouter

Interested in NER but have already realised the task is difficult? With large advancements in AI NER still remains a difficult task. Beyond having lots of labeled data few-shot learning can be used. To improve on the technique old ideas have come back to life. Modeling the labels as guassian distribution have shown to provide improvements in few shot NER models. #NER #machinelearning #ai #nlp #themlc.ca Das, Sarkar Snigdha Sarathi, Arzoo Katiyar, Rebecca J. Passonneau, and Rui Zhang. “CONTaiNER: Few-Shot Named Entity Recognition via Contrastive Learning.” arXiv, March 28, 2022. [https://lnkd.in/eUZGYFBd).

pub.towardsai.net: The article discusses the top open LLMs and encourages readers to explore further on Towards AI.

https://rebrand.ly/i378lo3 "Artificial intelligence (“AI”) is an outstanding technology for analyzing massive volumes of structured and unstructured data. Within the theoretical and methodological framework of Criminal Network Analysis (“CNA”), AI assists in classifying and extracting structured and unstructured data, and generating Criminal Network Graphs (“CNG”). To accomplish these tasks, Machine Learning Models must be trained with real/empirical data that describes the characteristics of the nodes/agents involved in the criminal networks and their interactions. This paper discusses the characteristics of the resulting models, herein defined as Machine Learning Models on Criminal Networks (MLMoCNs), and the prospects and obstacles for applying the most sophisticated AI techniques to CNA."

For my Library AI project: I wonder if is it possible: to have 2 trained LLMs, merge them, retrain them if needed, and obtain a new LLM?

“The pace of compute-intensive model releases is accelerating. Only 11 models exceeded 1023 FLOP in 2020. By 2024, this grew to 81 models in our dataset, and the trend shows no sign of slowing.” Wow! This is a quick but in depth review of landscape! Erik Schwartz great synopsis!

Looking to fine-tune an LLM for your specific use case? Why not make it faster and use less memory as well? In the recent AICamp gathering in Sydney about AI and LLMs, Daniel Han presented the Unsloth AI open-source framework for fine-tuning LLMs that promises up to 30x faster training and 60% less memory usage compared to traditional methods. The optimization is done by first optimizing all the mathematical equations used in the Language Model and then optimizing the implementation using Triton for better parallelism and GPU utilization. Check their GitHub page to see if it helps your task at hand: https://lnkd.in/gDfgme2W

Why do LLMs struggle with generating long-form content? Check out this paper to learn about the challenges LLMs face when tasked with producing extensive outputs, and how a change in framework addresses this limitation. By breaking down complex writing tasks and tuning models, this approach enables LLMs to generate coherent content exceeding 20,000 words. What are your thoughts? https://lnkd.in/eggWWnza #AI #LLM #Innovation

LLM context windows got huge. You can fit almost everything into the AI prompt. Almost. What to do when you have too much data? If you use Langchain, you may add a MapReduceChain to your code to handle documents longer than the context length of your AI model. It automatically splits the documents into smaller chunks, passes them to AI (with your prompt!), and combines the results. Link in the first comment

LLMs struggle to generate long outputs. This is sort of expected because these LLMs are rarely tuned on long-output examples (due to scarcity). This new paper proposes AgentWrite to enable off-the-shelf LLMs to generate coherent outputs beyond 20K words. AgentWrite breaks down the long generation task into subtasks and in a divide-and-conquer approach generates. The agent breaks the task into multiple writing subtasks and concatenates the outputs to get a final output (i.e., plan + write). The approach is then used to build SFT datasets that are used to tune LLMs to automatically generate coherent longer outputs. "Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models." https://lnkd.in/eggWWnza ↓ Join 75K+ AI researchers and devs so you don’t miss my weekly summary of the top AI and LLM papers: https://lnkd.in/e6ajg945