Dor Ovadia C.
๐ฎ๐ฑ Machine Learning Engineer & MLOps Specialist | M.Sc. ML Safety, Technion
- Report this post
Meta releases Llama Guard, a 7B parameter input-output safeguard model. It can be used to classify content in both LLM inputs and responses.Basically, it's an LLM that generates text indicating whether a prompt or response is ๐ฌ๐๐๐ or ๐ฎ๐ง๐ฌ๐๐๐, and if unsafe, it also lists the violating subcategories (example provided).--The model is released as part of the "Purple Llama" project by Meta:"๐๐ถ๐ณ๐ฑ๐ญ๐ฆ ๐๐ญ๐ข๐ฎ๐ข ๐ช๐ด ๐ข ๐ข๐ฏ ๐ถ๐ฎ๐ฃ๐ณ๐ฆ๐ญ๐ญ๐ข ๐ฑ๐ณ๐ฐ๐ซ๐ฆ๐ค๐ต ๐ต๐ฉ๐ข๐ต ๐ฐ๐ท๐ฆ๐ณ ๐ต๐ช๐ฎ๐ฆ ๐ธ๐ช๐ญ๐ญ ๐ฃ๐ณ๐ช๐ฏ๐จ ๐ต๐ฐ๐จ๐ฆ๐ต๐ฉ๐ฆ๐ณ ๐ต๐ฐ๐ฐ๐ญ๐ด ๐ข๐ฏ๐ฅ ๐ฆ๐ท๐ข๐ญ๐ด ๐ต๐ฐ ๐ฉ๐ฆ๐ญ๐ฑ ๐ต๐ฉ๐ฆ ๐ค๐ฐ๐ฎ๐ฎ๐ถ๐ฏ๐ช๐ต๐บ ๐ฃ๐ถ๐ช๐ญ๐ฅ ๐ณ๐ฆ๐ด๐ฑ๐ฐ๐ฏ๐ด๐ช๐ฃ๐ญ๐บ ๐ธ๐ช๐ต๐ฉ ๐ฐ๐ฑ๐ฆ๐ฏ ๐จ๐ฆ๐ฏ๐ฆ๐ณ๐ข๐ต๐ช๐ท๐ฆ ๐๐ ๐ฎ๐ฐ๐ฅ๐ฆ๐ญ๐ด. ๐๐ฉ๐ฆ ๐ช๐ฏ๐ช๐ต๐ช๐ข๐ญ ๐ณ๐ฆ๐ญ๐ฆ๐ข๐ด๐ฆ ๐ธ๐ช๐ญ๐ญ ๐ช๐ฏ๐ค๐ญ๐ถ๐ฅ๐ฆ ๐ต๐ฐ๐ฐ๐ญ๐ด ๐ข๐ฏ๐ฅ ๐ฆ๐ท๐ข๐ญ๐ด ๐ง๐ฐ๐ณ ๐๐บ๐ฃ๐ฆ๐ณ ๐๐ฆ๐ค๐ถ๐ณ๐ช๐ต๐บ ๐ข๐ฏ๐ฅ ๐๐ฏ๐ฑ๐ถ๐ต/๐๐ถ๐ต๐ฑ๐ถ๐ต ๐ด๐ข๐ง๐ฆ๐จ๐ถ๐ข๐ณ๐ฅ๐ด ๐ฃ๐ถ๐ต ๐ธ๐ฆ ๐ฑ๐ญ๐ข๐ฏ ๐ต๐ฐ ๐ค๐ฐ๐ฏ๐ต๐ณ๐ช๐ฃ๐ถ๐ต๐ฆ ๐ฎ๐ฐ๐ณ๐ฆ ๐ช๐ฏ ๐ต๐ฉ๐ฆ ๐ฏ๐ฆ๐ข๐ณ ๐ง๐ถ๐ต๐ถ๐ณ๐ฆ."--Have you heard of similar initiatives? Share your thoughts!#llm #mlsafety #aisafety
1
1 Comment
Dor Ovadia C.
๐ฎ๐ฑ Machine Learning Engineer & MLOps Specialist | M.Sc. ML Safety, Technion
10mo
- Report this comment
1Reaction
To view or add a comment, sign in
More Relevant Posts
-
Igor Jacques
Software Engineer at Hero Create | Javascript | Typescript | Kotlin | Golang | Verse | Microservices | Serverless | AWS
- Report this post
405 billion parameters โ Meta just launched their new version of Llama LLM, Llama 3.1, which includes a lot upgrades, but the most significant one is the Llama-405B Model, paraphrasing uncle Zuck's company: "The most capable openly available LLM to date".Meta has grown a lot in recent years by investing in high-quality open-source AI models. This update only shows the commitment of the company has with the area.We can also see that Meta is investing responsibly in security, so that developers build their solutions responsibly.Original post: https://go.fb.me/a24u0h
1
Like CommentSee AlsoWhat is a GREAT Roblox Game? | Mateen Soudagar posted on the topic | LinkedInThe PS5 is on sale for Prime Big Deal Days. Now's the time to buy | Digital TrendsWhat you need to know this week in #AI | Galit Galperin posted on the topic | LinkedInCalgary Flames seek stability, progress and playoffs with younger rosterTo view or add a comment, sign in
-
Devilal Sharma
Co-founder, CEO - Famdo.Health | IIT Madras
- Report this post
๐จ๐ป๐น๐ฒ๐ฎ๐๐ต๐ถ๐ป๐ด ๐๐ต๐ฒ ๐ฃ๐ผ๐๐ฒ๐ฟ ๐ผ๐ณ ๐๐น๐ฎ๐บ๐ฎ ๐ฏ.๐ฎ: ๐ ๐๐ฎ๐บ๐ฒ-๐๐ต๐ฎ๐ป๐ด๐ฒ๐ฟ ๐ถ๐ป ๐๐ฒ๐ป๐๐!If you're not amazed, what will it take? Meta has just released Llama 3.2, and I'm blown away! Watch the video below to experience it for yourself.Llama 3.2 boasts best-in-class vision and image capabilities, revolutionizing image reasoning use cases. Key highlights include:โ โ Smaller models for edge use cases: Lightweight 1B and 3B models deliver impressive performance.โ Instantaneous prompts and responses.โ โ Local model execution ensures data privacy.โ Open-source, empowering developers to innovate.I'm particularly excited about the small model capabilities - perfect for mobile apps and edge computing use cases. This update is a significant tailwind!Check it out and discover the endless possibilities! #Llama32 #AI #GenAI #OpenSource #EdgeComputing #innovation #gaming
49
3 Comments
Like CommentTo view or add a comment, sign in
-
Chandra Sekhar Medisetti
Gen AI | Python | Conversational AI | NLP | RASA | LLMs | AI | open-source
- Report this post
๐ Exciting News! Meta has unveiled Llama3, introducing a host of benefits for users. What's even better? These models are open-source, ensuring accessibility for all. Whether you're keen to explore them via Meta's website or through Hugging Face, the possibilities are endless.๐ Here are some key highlights:โ Available in 8B and 70B pretrained and instruction-tuned versions.โ Llama 3 models expand data scaling capabilities.โ Trained on two 24K GPU clusters.โ Utilizes over 15T tokens, marking a 7x increase from Llama 2's dataset size.โ Dataset encompasses 4x more code than Llama.โ Supports an impressive 8K context length, doubling Llama 2's capacity.And that's not all! For Windows users, rejoiceโOllama is supported locally.website: https://ai.meta.com/๐ก Don't miss out on the opportunity to leverage these powerful tools for your projects. Dive in, explore, and innovate with Llama3 today! #Meta #Llama3 #OpenSource #Innovation #Technology#conversational
3
Like CommentTo view or add a comment, sign in
-
Junaid Syed
Data Scientist at SLB | MS Analytics Georgia Tech | Artificial Intelligence | Machine Learning
- Report this post
๐ Llama 3.2: Meta's Latest Breakthrough! ๐ Meta has just unleashed Llama 3.2, a versatile collection of large language models (LLMs). From lightweight text-only models for mobile devices to powerful models for image reasoning, this release has it all. Plus, the new Llama Stack makes development with Llama models a breeze. Key Features Unpacked - Edge and Mobile Capabilities: Lightweight, text-only models (1B and 3B parameters) designed for on-device tasks like summarization, rewriting, and instruction following. - Vision Capabilities: Larger models (11B and 90B) excel in enabling image captioning, document understanding, and visual grounding tasks. - Llama Stack: A standardized interface making development and deployment of Llama-based applications across different environments seamless. - Enhanced Safety: Updated safety measures along with Llama Guard 3 for both text and image input filtering. The Perks of Local Deployment Running Llama 3.2 locally offers two key advantages: - Speed: Faster response times with on-device processing. - Privacy: Keep your data secure and under your control. Ready to dive in? Download Llama 3.2 now from the official Llama website or Hugging Face.
Like CommentTo view or add a comment, sign in
-
Mateen Soudagar
CEO @ Metakey | ๐ฅ๐ผ๐ฏ๐น๐ผ๐ /๐๐ผ๐ฟ๐๐ป๐ถ๐๐ฒ/๐ ๐ฒ๐๐ฎ๐๐ฒ๐ฟ๐๐ฒ ๐ฆ๐๐ฟ๐ฎ๐๐ฒ๐ด๐ | KPMG, Cap Gemini, Aus Open, One Time Square, 30+ clients | DMs Open.
- Report this post
Meta is fighting AI images big time by thinking of watermarking all web images.They're doing this to fight against fake images. They're not only marking AI-made images on their own apps, (Facebook, Instagram, etc),but they're also going to label AI-created images from other services (Midjourney, Google, OpoenAI, Microsoft, Shutterstock, etc).These AI images have hidden signs like extra information or secret marks that show they were made by a computer program.The head of Global Affairs at Meta, Nick Clegg, saysโIt's important to be able to tell if an image was made by a human or a computer.โIn the future, Meta will also label AI-made audio and video content. They're even trying to develop new technology to find AI-made content without any obvious marks.This is important to stop the spread of fake videos and images made by AI, and to make sure we can trust what we see online.โ Building Roblox games atthemetakey.comโ Follow Mateen Soudagar for more insights into#web3x#gamingx#brands#marketing#fortnite#genz
16
Like CommentTo view or add a comment, sign in
-
-
Damien Orozco
Damien Orozco is an Influencer
Founder & Chief Helpfulness Officer at Xocial Gong. Your personal Link(Ed)In Net(Work)ER + advocate, eager to help fill your calendar with your ideal clients. Let's grow your business together ๐ TODAY! ๐๐ฝ 323-405-3872
- Report this post
On December 6th, META (a.k.a. Facebook) introduced โก๏ธ image.meta.com Have you tried it yet? Did you even know they launched it? If not, now you know. Check out one of my first creations using the following prompt 'a cyborg spider hanging on a web'. I think it's super awesome. So if you are struggling with getting images for your content, this may be the answer to your struggle ๐๐ค๐ฅณ๐๐๐พ#imaginewithmeta #cyborgspider #GenerativeAI #generativeart
Like CommentTo view or add a comment, sign in
-
Xocial Gong, LLC
313 followers
- Report this post
On December 6th, META (a.k.a. Facebook) introduced โก๏ธ image.meta.com Have you tried it yet? Did you even know they launched it? If not, now you know. Check out one of my first creations using the following prompt 'a cyborg spider hanging on a web'. I think it's super awesome. So if you are struggling with getting images for your content, this may be the answer to your struggle ๐๐ค๐ฅณ๐๐๐พ#imaginewithmeta #cyborgspider #GenerativeAI #generativeart
Like CommentTo view or add a comment, sign in
-
Dan C.
Digital Marketer. Has more insight than LensCrafters. High Purview with a cost per view to match. โ๐ป๏ธ
- Report this post
"Why do companies like Meta release their local LLMs for free?"This question came up on reddit and the answers are spot on. This also applies to other open source LLM's in market. "If you make LLMs based on public data ubiquitous quickly, then private data - which FB has a lot of - becomes more valuable. It reduces the value of OpenAI and increases FB's own valuation in the long run. They have a lot of multimodal data (pics, video and videogame/metaverse) that will add value to the next generation models. When that transition happens in the next few years they will start holding back on the models."Calling it now; we will see a paid subscription offering for Meta AI by the end of the decade. #CryInCapitalGains โLink: https://lnkd.in/eRpWmhiR
5
Like CommentTo view or add a comment, sign in
-
Igor Barshteyn
๐ Protecting Data and Mitigating Information Security and Privacy Risks *All Opinions Are Strictly My Own *All My Words Are Human-Generated *Don't Sell Me Things!
- Report this post
The first (small) versions of LLaMA-3 are coming as early as next week from Meta.https://lnkd.in/eRM65n43Hopes are high that Meta's release is going to define the new state-of-the-art for local large language models. We shall see. The competition is fierce. Currently, the "smartest" local model is Command-R+ by Cohere (https://lnkd.in/eDhnwHCw, can be downloaded in GGUF formats here https://lnkd.in/eMp47845 and can be run on llama.cpp using the PR here https://lnkd.in/eateBYA2) but it's 104B parameters, and is rather unwieldy in terms of resource use.Fingers crossed for Meta. The last 2 times, Meta did not dissappoint.
4
2 Comments
Like CommentTo view or add a comment, sign in
1,590 followers
- 112 Posts
View Profile
FollowExplore topics
- Sales
- Marketing
- IT Services
- Business Administration
- HR Management
- Engineering
- Soft Skills
- See All