6 ms·
Beginner's Guide to Llama Models
- speedgoose 3y agoWhat are the hardware requirements to fine tune llama ?
- bart__ 3y ago~20 GB vram for the 7B model and 48 GB for the 13B model. It depends on the context size as well. I'd recommend renting a 4090 from a cloud provider like runpod/vast ai to get started, using a PEFT tutorial.
- speedgoose 3y agoThanks. What about the 70B model? I assume a 4090 will not be enough. Is it linear system requirements ?
- bart__ 3y ago4090 only has 24 GB and will only be able to fine tune (and merge, which is more memory intensive) the 7B model. The RTX6000 with 48 GB is able to fine tune the 13B model. The 70B model presumably needs multiple GPUs, like 4 RTX6000. For people starting out, you can also use a free GPU from Google colab to fine tune a 7B model. Finetuning 70B gets more expensive and I would suggest trying smaller models first with a high quality dataset. It is mostly linear I think.
- speedgoose 3y agoThanks. My plan is to use this research cluster: https://www.ex3.simula.no/resources https://www.ex3.simula.no/resources I will probably train how to fine tune on the small model but I don’t really need to use a worse model to save money.
- cowthulhu 3y agoDoes anyone know how well a fine-tuned Llama model will do compared to GPT 4 on complex tasks?
- thatguymike 3y agoI find the whole site being covered in images of skinny waifu girls... offputting. The guide itself seems fine, if high-level. [This](https://chat.lmsys.org/?leaderboard https://chat.lmsys.org/?leaderboard) is a really nice link I hadn't seen before.
- mufti_menk 3y agoI found them beautiful
- cypress66 3y agoYeah, a lot of them are very cool.
- brabel 3y agoSame here... the images are beautiful but seem completely unrelated to the topic (other than being AI-generated - but they could've generated images that have some sort of relation to the topic instead!). It kind of shows what the author has been using the AI for, I suppose :P.
- kristofferg 3y agoI would usually agree but in this case its a pun on “models”. I find it quite funny.
- brabel 3y agoOh I see... and a "model" is always young, skinny, white and female I guess (cannot blame the AI though, I guess that is what you get if you look at the internet as a whole as your training data). EDIT: I had to lookup "waifu" and it seems the author probably prompted for "waifu model"(?)... according to Wikipedia, for those like me who are not into that sort of thing: "A Waifu is an illustrated female character from an anime or any non-live action media in which an individual becomes sexually attracted to."
- MahdeenSky 3y ago+1 uncensored models
- neilv 3y ago> How to install Llama models? > See the installation guide for Windows and the installation guide for Mac. Much "open" LLM/SD/etc. grassroots stuff seems to be shooting themselves in the face, by pushing others to closed platforms. Now is one of the times to be increasing pressure for open platforms, not backsliding.
- abwizz 3y agoyes, they are, but it seems a good fit nonetheless
- milar 3y agoThis is literally AI generated txt and gen’d images. HN isn’t prepared yet. Prepare for 100x more of these.
- raincole 3y agoChecked the submitter's history. Clearly "spam my content farm posts on HN and hope one of them gets lucky".
- yu3zhou4 3y agoCan you recommend another systematized guides like this about LLMs and ML in general? Even though was quite limited in info, I like the structured and concise form and I’d be happy to learn about similar blog posts
- jmorgan 3y agoI really enjoyed Anrej Kaparthy's llama2.c project (https://github.com/karpathy/llama2.c https://github.com/karpathy/llama2.c), which runs through creating and running a miniature Llama2 architecture model from scratch.
- jmorgan 3y agoAn easy way to try many of the fine-tuned Llama 2 models is https://github.com/jmorganca/ollama https://github.com/jmorganca/ollama. A maintainer of the project has been collecting a full list here (with different quantization levels), most of which are Llama 2-based: https://gist.github.com/mchiang0610/b959e3c189ec1e948e4f6a1f737a1fc5 https://gist.github.com/mchiang0610/b959e3c189ec1e948e4f6a1f... Since the release of Llama 2 the number of models based on it has been growing significantly.. some popular ones: - codeup (A code generation model - DeepSE) - llama2-uncensored (George Sung) - nous-hermes-llama2 (Nous Research) - wizardlm-uncensored (WizardLM) - stablebeluga (Stability AI) The article also recommends oobabooga's text-generation-webui which includes a full web dashboard.
- danielbln 3y agollama2-uncensored is a lot of fun, interacting with an unaligned/uncensored model feels very fresh.
- xcdzvyn 3y agoWow, thanks! I really like Ollama, so I'm glad the models listed at the top of the Readme aren't all you could use. Do you know where the data in that Gist came from? Is there a registry somewhere?
- jmorgan 3y agoThere is! While not easy to use yet, there's a sort-of-hidden way models can be listed with: curl https://ollama.ai/v2/_catalog | jq Then to list "tags" for a given model (e.g. llama2): curl https://ollama.ai/v2/library/llama2/tags/list | jq
- xcdzvyn 3y agoOh, you're the maintainer! Hahaha my bad. Thanks :)
- ulnarkressty 3y agoI tried multiple flavors of llama models, they are all quite dumb. Even the 70b parameter one. It knows about more things which the smaller models just hallucinate when asked, but still cannot do even slightly more complex tasks. I'm also not sure about the current testing methodologies i.e. the 'passed the SAT' hype. Given that the training set already contains much of the information, we should probably compare the AI results with humans having unlimited time and access to the required material.
- smcin 3y agoPost us some sample prompts and answers please.
- avion23 3y agoI've evaluated some models. The best ones are based on llama-2. Good is for example codeup-llama-2-13b-chat-hf Uncensored (partially) is nous-hermes-llama2-13b
- marcopicentini 3y agoCan I train a llama 2 model on custom data and make it expert on my data knowledge? For example, If I give it thousands of law pages it will be a domain expert about law ?
- jbjbjbjb 3y agoI tried running a llama model using oobabooga but kept running into one problem after the next. Anyone know of a config that might work (even quite slowly) on an i9 laptop, 32gb ram with nvidia graphics 8gb?
- crossroadsguy 3y agoWhat's a good and friendly place to start for a seasoned Android developer to get a peek into the world of AI/ML as we are seeing all this today (or specifically last few months maybe), who is kind of having a FOMO cum curiosity (and a little worry about career/future and all) about all this? To get a taste of things.
- tony12345678 3y agoAs a Mac user, I think Ollama is amazing. Thank you! :). Is there any chance that functionality could be added for fine-tuning (e.g. document/text file uploads)?