Meta Llama 3(llama.meta.com)
llama.meta.com
Meta Llama 3
https://llama.meta.com/llama3/
923 comments
They've got a console for it as well,
https://www.meta.ai/
And announcing a lot of integration across the Meta product suite, https://about.fb.com/news/2024/04/meta-ai-assistant-built-wi...
Neglected to include comparisons against GPT-4-Turbo or Claude Opus, so I guess it's far from being a frontier model. We'll see how it fares in the LLM Arena.
And announcing a lot of integration across the Meta product suite, https://about.fb.com/news/2024/04/meta-ai-assistant-built-wi...
Neglected to include comparisons against GPT-4-Turbo or Claude Opus, so I guess it's far from being a frontier model. We'll see how it fares in the LLM Arena.
Public benchmarks are broadly indicative, but devs really should run custom benchmarks on their own use cases.
Replicate created a Llama 3 API [0] very quickly. This can be used to run simple benchmarks with promptfoo [1] comparing Llama 3 vs Mixtral, GPT, Claude, and others:
Edit: ollama now supports Llama 3 8b, making it easy to run this eval locally.
[1] https://github.com/typpo/promptfoo
Replicate created a Llama 3 API [0] very quickly. This can be used to run simple benchmarks with promptfoo [1] comparing Llama 3 vs Mixtral, GPT, Claude, and others:
prompts:
- 'Answer this programming question concisely: {{ask}}'
providers:
- replicate:meta/meta-llama-3-8b-instruct
- replicate:meta/meta-llama-3-70b-instruct
- replicate:mistralai/mixtral-8x7b-instruct-v0.1
- openai:chat:gpt-4-turbo
- anthropic:messages:claude-3-opus-20240229
tests:
- vars:
ask: Return the nth element of the Fibonacci sequence
- vars:
ask: Write pong in HTML
# ...
Edit: ollama now supports Llama 3 8b, making it easy to run this eval locally.
providers:
- ollama:chat:llama3
[1] https://github.com/typpo/promptfoo
Llama 3 70B has debuted on the famous LMSYS chatbot arena leaderboard at position number 5, tied with Claude 2 Sonnet, Bard (Gemini Pro), and Command R+, ahead of Claude 2 Haiku and older versions of GPT-4.
The score still has a large uncertainty so it will take a while to determine the exact ranking and things may change.
Llama 3 8B is at #12 tied with Claude 1, Mixtral 8x22B, and Qwen-1.5-72B.
These rankings seem very impressive to me, on the most trusted benchmark around! Check the latest updates at https://arena.lmsys.org/
Edit: On the English-only leaderboard Llama 3 70B is doing even better, hovering at the very top with GPT-4 and Claude Opus. Very impressive! People seem to be saying that Llama 3's safety tuning is much less severe than before so my speculation is that this is due to reduced refusal of prompts more than increased knowledge or reasoning, given the eval scores. But still, a real and useful improvement! At this rate, the 400B is practically guaranteed to dominate.
The score still has a large uncertainty so it will take a while to determine the exact ranking and things may change.
Llama 3 8B is at #12 tied with Claude 1, Mixtral 8x22B, and Qwen-1.5-72B.
These rankings seem very impressive to me, on the most trusted benchmark around! Check the latest updates at https://arena.lmsys.org/
Edit: On the English-only leaderboard Llama 3 70B is doing even better, hovering at the very top with GPT-4 and Claude Opus. Very impressive! People seem to be saying that Llama 3's safety tuning is much less severe than before so my speculation is that this is due to reduced refusal of prompts more than increased knowledge or reasoning, given the eval scores. But still, a real and useful improvement! At this rate, the 400B is practically guaranteed to dominate.
I tried generating a Chinese rap song, and it did generate a pretty good rap. However, upon completion, it deleted the response, and showed
> I don’t understand Chinese yet, but I’m working on it. I will send you a message when we can talk in Chinese.
I tried some other languages and the same. It will generate non-English language, but once its done, the response is deleted and replaced with the message
I tried some other languages and the same. It will generate non-English language, but once its done, the response is deleted and replaced with the message
Lots of great details in the blog: https://ai.meta.com/blog/meta-llama-3/
Looks like there's a 400B version coming up that will be much better than GPT-4 and Claude Opus too. Decentralization and OSS for the win!
Looks like there's a 400B version coming up that will be much better than GPT-4 and Claude Opus too. Decentralization and OSS for the win!
I just want to express how grateful I am that Zuck and Yann and the rest of the Meta team have adopted an open approach and are sharing the model weights, the tokenizer, information about the training data, etc. They, more than anyone else, are responsible for the explosion of open research and improvement that has happened with things like llama.cpp that now allow you to run quite decent models locally on consumer hardware in a way that you can avoid any censorship or controls.
Not that I even want to make inference requests that would run afoul of the controls put in place by OpenAI and Anthropic (I mostly use it for coding stuff), but I hate the idea of this powerful technology being behind walls and having gate-keepers controlling how you can use it.
Obviously, there are plenty of people and companies out there that also believe in the open approach. But they don't have hundreds of billions of dollars of capital and billions in sustainable annual cash flow and literally ten(s) of billions of dollars worth of GPUs! So it's a lot more impactful when they do it. And it basically sets the ground rules for everyone else, so that Mistral now also feels compelled to release model weights for most of their models.
Anyway, Zuck didn't have to go this way. If Facebook were run by "professional" outside managers of the HBS/McKinsey ilk, I think it's quite unlikely that they would be this open with everything, especially after investing so much capital and energy into it. But I am very grateful that they are, and think we all benefit hugely from not only their willingness to be open and share, but also to not use pessimistic AI "doomerism" as an excuse to hide the crown jewels and put it behind a centralized API with a gatekeeper because of "AI safety risks." Thanks Zuck!
Not that I even want to make inference requests that would run afoul of the controls put in place by OpenAI and Anthropic (I mostly use it for coding stuff), but I hate the idea of this powerful technology being behind walls and having gate-keepers controlling how you can use it.
Obviously, there are plenty of people and companies out there that also believe in the open approach. But they don't have hundreds of billions of dollars of capital and billions in sustainable annual cash flow and literally ten(s) of billions of dollars worth of GPUs! So it's a lot more impactful when they do it. And it basically sets the ground rules for everyone else, so that Mistral now also feels compelled to release model weights for most of their models.
Anyway, Zuck didn't have to go this way. If Facebook were run by "professional" outside managers of the HBS/McKinsey ilk, I think it's quite unlikely that they would be this open with everything, especially after investing so much capital and energy into it. But I am very grateful that they are, and think we all benefit hugely from not only their willingness to be open and share, but also to not use pessimistic AI "doomerism" as an excuse to hide the crown jewels and put it behind a centralized API with a gatekeeper because of "AI safety risks." Thanks Zuck!
I was curious how the numbers compare to GPT-4 in the paid ChatGPT Plus, since they don't compare directly themselves.
GPT-4 numbers from from https://github.com/openai/simple-evals gpt-4-turbo-2024-04-09 (chatgpt)
Llama 3 8B Llama 3 70B GPT-4
MMLU 68.4 82.0 86.5
GPQA 34.2 39.5 49.1
MATH 30.0 50.4 72.2
HumanEval 62.2 81.7 87.6
DROP 58.4 79.7 85.4
GPT-4 numbers from from https://github.com/openai/simple-evals gpt-4-turbo-2024-04-09 (chatgpt)
Zuck has an interview out for it as well,
https://twitter.com/dwarkesh_sp/status/1780990840179187715
The model card has the benchmark results relative to other Llama models including Llama 2: https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md...
The dramatic performance increase of Llama 3 relative to Llama 2 (even Llama 2 13B!) is very impressive. Doubling the context window to 8k will open a lot of new oppertunities too.
The dramatic performance increase of Llama 3 relative to Llama 2 (even Llama 2 13B!) is very impressive. Doubling the context window to 8k will open a lot of new oppertunities too.
https://github.com/meta-llama/llama3/blob/main/LICENSE
Llama is not open source. It's corporate freeware with some generous allowances.
Open source licenses are a well defined thing. Meta marketing saying otherwise doesn't mean they get to usurp the meaning of a well understood and commonly used understanding of the term "open source."
https://opensource.org/license
Nothing about Meta's license is open source. It's a carefully constructed legal agreement intended to prevent any meaningful encroachment by anyone, ever, into any potential Meta profit, and to disavow liability to prevent reputational harm in the case of someone using their freeware for something embarrassing.
If you use it against the license anyway, you'll just have to hope you never get successful enough that it becomes more profitable to sue you and take your product away than it would be annoying to prosecute you under their legal rights. When the threshold between annoying and profitable is crossed, Meta's lawyers will start sniping and acquiring users of their IP.
Llama is not open source. It's corporate freeware with some generous allowances.
Open source licenses are a well defined thing. Meta marketing saying otherwise doesn't mean they get to usurp the meaning of a well understood and commonly used understanding of the term "open source."
https://opensource.org/license
Nothing about Meta's license is open source. It's a carefully constructed legal agreement intended to prevent any meaningful encroachment by anyone, ever, into any potential Meta profit, and to disavow liability to prevent reputational harm in the case of someone using their freeware for something embarrassing.
If you use it against the license anyway, you'll just have to hope you never get successful enough that it becomes more profitable to sue you and take your product away than it would be annoying to prosecute you under their legal rights. When the threshold between annoying and profitable is crossed, Meta's lawyers will start sniping and acquiring users of their IP.
Just got uploaded to HuggingFace: https://huggingface.co/meta-llama/Meta-Llama-3-8B https://huggingface.co/meta-llama/Meta-Llama-3-70B
I am always excited to see these Open Weight models released, I think its very good for the ecosystem and definitely has its place in many situations.
However since I use LLMs as a coding assistant (mostly via "rubber duck" debugging and new library exploration) I really don't want to use anything other than the absolutely best in class available now. That continues to be GPT4-turbo (or maybe Claude 3).
Does anyone know if there is any model out there that can be run locally and compete with GPT4-turbo? Or am I asking for something that is impossible?
However since I use LLMs as a coding assistant (mostly via "rubber duck" debugging and new library exploration) I really don't want to use anything other than the absolutely best in class available now. That continues to be GPT4-turbo (or maybe Claude 3).
Does anyone know if there is any model out there that can be run locally and compete with GPT4-turbo? Or am I asking for something that is impossible?
From the article
>We made several new observations on scaling behavior during the development of Llama 3. For example, while the Chinchilla-optimal amount of training compute for an 8B parameter model corresponds to ~200B tokens, we found that model performance continues to improve even after the model is trained on two orders of magnitude more data. Both our 8B and 70B parameter models continued to improve log-linearly after we trained them on up to 15T tokens. Larger models can match the performance of these smaller models with less training compute, but smaller models are generally preferred because they are much more efficient during inference.
Can someone experienced please explain this. Does this mean, a lean model with more training time and/or more (or better) training data will perform better than a fat model?
>We made several new observations on scaling behavior during the development of Llama 3. For example, while the Chinchilla-optimal amount of training compute for an 8B parameter model corresponds to ~200B tokens, we found that model performance continues to improve even after the model is trained on two orders of magnitude more data. Both our 8B and 70B parameter models continued to improve log-linearly after we trained them on up to 15T tokens. Larger models can match the performance of these smaller models with less training compute, but smaller models are generally preferred because they are much more efficient during inference.
Can someone experienced please explain this. Does this mean, a lean model with more training time and/or more (or better) training data will perform better than a fat model?
I'm a big fan of various AI companies taking different approaches. OpenAI keeping it close to their hearts but have great developer apis. Meta and Mistral going open weights + open code. Anthropic and Claude doing their thing.
Competition is a beautiful thing.
I am half excited and half scared that AGI is our generation's space war.
I hope we can solve the big human problems, instead of more scammy ads and videos.
So far AI has been more hype than substance.
Competition is a beautiful thing.
I am half excited and half scared that AGI is our generation's space war.
I hope we can solve the big human problems, instead of more scammy ads and videos.
So far AI has been more hype than substance.
"You’ll also soon be able to test multimodal Meta AI on our Ray-Ban Meta smart glasses."
Now this is interesting. I've been thinking for some time now that traditional computer/smartphone interfaces are on the way out for all but a few niche applications.
Instead, everyone will have their own AI assistant, which you'll interact with naturally the same way as you interact with other people. Need something visual? Just ask for the latest stock graph for MSFT for example.
We'll still need traditional interfaces for some things like programming, industrial control systems etc...
Now this is interesting. I've been thinking for some time now that traditional computer/smartphone interfaces are on the way out for all but a few niche applications.
Instead, everyone will have their own AI assistant, which you'll interact with naturally the same way as you interact with other people. Need something visual? Just ask for the latest stock graph for MSFT for example.
We'll still need traditional interfaces for some things like programming, industrial control systems etc...
Quick thoughts -
Major arch changes are not that major, mostly GQA and tokenizer improvements. Tokenizer improvement is a under-explored domain IMO.
15T tokens is a ton!
400B model performance looks great, can’t wait for that to be released. Might be time to invest in a Mac studio!
OpenAI probably needs to release GPT-5 soon to convince people they are still staying ahead.
Major arch changes are not that major, mostly GQA and tokenizer improvements. Tokenizer improvement is a under-explored domain IMO.
15T tokens is a ton!
400B model performance looks great, can’t wait for that to be released. Might be time to invest in a Mac studio!
OpenAI probably needs to release GPT-5 soon to convince people they are still staying ahead.
Interesting, the 8B model was trained for 1.3M hours, while the 70B for 6.4M hours at 700W. Assuming $0.05/kWh (WA price) it's $46k and $224k. Even allowing for cooling, CPUs, and more expensive power wherever they are running this, still well less than $1M in power. I somehow thought it would be much more.
The nVidia bill is another matter - assuming 5 year amortization and $45k H100, it works out $1/h, so $8M or so.
The nVidia bill is another matter - assuming 5 year amortization and $45k H100, it works out $1/h, so $8M or so.
Just a quick observation: it seems to not mention commercial companies (or at least be biased against it). I tried executing "what are popular design tools with an infinite canvas" against both meta.ai and OpenAI. OpenAI returned what you would expect, Figma Sketch etc. But MetaAI only returned free/open-source software https://x.com/seveibar/status/1781042926430437404
The amount of open source stuff Facebook (Meta) has given us over the years is astounding. pytorch, dall-e, react, react native, graphql, cassandra, tor. Commoditized VR, love my quest 3. Just an incredible track record. We're lucky they release all this stuff for free. The Zuck is one of a kind.
If any one is interesting in seeing how 400B model compares with other opensource models, here is a useful chart: https://x.com/natolambert/status/1780993655274414123
I'm so surprised that Meta is actually leading the open source AI landscape?! I've used llama2 extensively and can't wait to try out llama3 now. I can't believe that it does better than Claude 3 in benchmarks (though admittedly claude 3 seems to have been nerfed recently)
I sure do wish there was more info about how its trained and its training data.
I sure do wish there was more info about how its trained and its training data.
Slightly off-topic, but I'd like to see a model that can fit nicely on a consumer 24GB GPU, such as a 20B model (8bit quantized, but some spare space to allow for context), or perhaps 40B so that a 4-bit quant could fit into 20GB VRAM. It's disappointing to me that Meta stopped releasing the 30B after llama 1. (I know codellama exists, but that's been finetuned to one use case).
Would love to experiment with this for work, but the following clause in the license (notably absent in the Llama 2 license) would make this really hard:
> i. If you distribute or make available the Llama Materials (or any derivative works thereof), or a product or service that uses any of them, including another AI model, you shall (A) provide a copy of this Agreement with any such Llama Materials; and (B) prominently display “Built with Meta Llama 3” on a related website, user interface, blogpost, about page, or product documentation. If you use the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include “Llama 3” at the beginning of any such AI model name.
Really impressive HumanEval results for the 8B model though, would love to plug this into Continue for tab completion since the current benchmark numbers for Llama 3 8B blow every other 8B model out of the water
> i. If you distribute or make available the Llama Materials (or any derivative works thereof), or a product or service that uses any of them, including another AI model, you shall (A) provide a copy of this Agreement with any such Llama Materials; and (B) prominently display “Built with Meta Llama 3” on a related website, user interface, blogpost, about page, or product documentation. If you use the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available, you shall also include “Llama 3” at the beginning of any such AI model name.
Really impressive HumanEval results for the 8B model though, would love to plug this into Continue for tab completion since the current benchmark numbers for Llama 3 8B blow every other 8B model out of the water
Surprisingly, looks like Mark Zuckerberg is listed as a contributor in the Model Card [1]. I thought since its a pretty big effort, most executives would be added to it as well, but that does not seem to be the case at all. In fact I was surprised that Soumith Chintala was left out here [2].
[1] - https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md...
[2] - https://soumith.ch/about/
[1] - https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md...
[2] - https://soumith.ch/about/
Open weight models do more for AI safety than any other measure by far, as the most serious threath is never going to be misuse, but abuse of unequal access.
What's the reason for the explosion of GGUF model uploads on Huggingface?
https://huggingface.co/models?sort=trending&search=llama-3
Sounds like quite the duplication of the model conversion. Few months ago it was TheBloke that did reliable model uploads. Nowadays, it's more straightforward to use https://ollama.com/library/llama3 as source to save time on searching through HF.
https://huggingface.co/models?sort=trending&search=llama-3
Sounds like quite the duplication of the model conversion. Few months ago it was TheBloke that did reliable model uploads. Nowadays, it's more straightforward to use https://ollama.com/library/llama3 as source to save time on searching through HF.
Last week, I made the explicit decision to circle through paid subscriptions of various providers every few months, instead of sticking to just ChatGPT.
Glad to see Meta.ai on the list, in addition to Claude, Gemini, Mistral and Perplexity.
Glad to see Meta.ai on the list, in addition to Claude, Gemini, Mistral and Perplexity.
I don't see any explanation for why they trained 8B instead of 7B.
I thought that If you have a 16GB GPU, you can put 14GB(7B*16bits) model into it, but how does it fit If the model is exactly 16GB?
Is there a download link for this model like LLAMA2 or is it going to be exclusively owned and operated by Meta this time?
and https://about.fb.com/news/2024/04/meta-ai-assistant-built-wi...
edit: and https://twitter.com/karpathy/status/1781028605709234613