Sanghamitra Deb

This talk on OLMo is must-watch. Covers each part of the LLM pipeline in detail, and the work of dozens of researchers to build, tune, and open-source it. (COLMo?) youtu.be/qMTzor0j418

The best way to open people's minds isn't to argue with them. It's to listen to them. When people feel understood, they become less defensive and more reflective—and develop less extreme, more nuanced views. Productive disagreements begin with curiosity, not persuasion.

In order to make sure your LLM is not biased, you need a good data discovery & measurement platform that ensures known biases do not exist in the training data. #LLM #guardrails #data @_odsc

The amount of literature review to understand evaluations is overwhelming! #llms #evaluations ⁦@_odsc⁩ ⁦@huggingface⁩

Security by obscurity! Building an LLM chat bot on enterprise data can lead exposure of sensitive data which were not discoverable before the chat bot. Enterprise can use #llms to detect sensitive info and decide to surface it. #trustandsafety #guardrails #moderation. #prompts.

A week ago, Meta presented Chain of Verification (arxiv.org/abs/2309.11495), that addresses and mitigates hallucination. It is a specific implementation of Reflexion, similar to DRA. I have added it to my hallucination guide #advancedprompting" target="_blank" rel="nofollow noopener">amatriain.net/blog/hallucina…

🔥🔥🔥 Introduce the newest WizardCoder 34B based on Code Llama. ✅WizardCoder-34B surpasses GPT-4, ChatGPT-3.5 and Claude-2 on HumanEval with 73.2% pass@1 🖥️Demo: http://47.103.63.15:50085/ 🏇Model Weights: huggingface.co/WizardLM/Wizar… 🏇Github: github.com/nlpxucan/Wizar… The 13B/7B versions are coming soon. *Note: There are two HumanEval results of GPT4 and ChatGPT-3.5: 1. The 67.0 and 48.1 are reported by the official GPT4 Report (2023/03/15) of OpenAI. 2. The 82.0 and 72.5 are tested by ourselves with the latest API (2023/08/26).

If you found this thread helpful and would like to learn a streamlined process to write papers that get published, this free training will be ideal! 👇✨ annaclemens.com/training?utm_s…

#9: Not asking for writing training 💰 If you feel under-skilled at writing, ask your PI, department or grad school for funding to enrol in a scientific writing course. Many universities have funding available for professional development - you just have to make the ask!

#3: Trying to make your sentences perfect in your first draft 🤓 If you keep deleting what you have written, you won’t get a lot of text on the page. The goal isn’t to get every sentence, every word right from the start but to refine your paper through editing.

Mitosis of Congress

Want to fine-tune Llama2 for your use case on your data? ICYMI—@scale_ai open-sourced LLM Engine, the easiest way to fine-tune and serve Llama2 and other open-source models See 🧵👇

Yay, llama2.c can now load and inference the Meta released models! :) E.g. here inferencing the smallest 7B model at ~3 tokens/s on 96 OMP threads on a cloud Linux box. Still just CPU, fp32, one single .c file of 500 lines: github.com/karpathy/llama… expecting ~300 tok/s tomorrow :)

Building a streaming chat agent in 1/2 hour with vercel AI. #generativeAI #frontend #llm

One year road maps are not the best technology practice in the space of #generativeAI #LLM #multimodal. @pinecone panel discussion. Agree from the past 6 months of developing LLm based services in the age of disruption.

Building a Multi-User Chatbot with Langchain and Pinecone in Next.JS | Pinecone pinecone.io/learn/javascri…

#opstack. Source data , chunk them and embed them & upload to pinecone. Someone comes in to chat. Application embeds the query and query against pinecone, contextual & semantic info. Send all this to openAI and get the answer correctly #generativeAI #llm

From Deep to Long Learning? · Hazy Research hazyresearch.stanford.edu/blog/2023-03-2…

For #GenerativeAI solutions, there are several options (1) building your own models (2) fine tuning (3) keeping information in vector databases to provide context to the #llm to reduce hallucinations. @pinecone

Vector databases help with #llm hallucinations by retrieving additional relevant , domain specific information to LLMs. #generativeai ⁦@pinecone⁩