mup
gpt-3
mup | gpt-3 | |
---|---|---|
12 | 41 | |
1,186 | 9,406 | |
3.4% | - | |
2.7 | 3.5 | |
7 days ago | over 3 years ago | |
Jupyter Notebook | ||
MIT License | - |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
mup
-
Announcing xAI July 12th 2023
Our team is led by Elon Musk, CEO of Tesla and SpaceX. We have previously worked at DeepMind, OpenAI, Google Research, Microsoft Research, Tesla, and the University of Toronto. Collectively we contributed some of the most widely used methods in the field, in particular the Adam optimizer, Batch Normalization, Layer Normalization, and the discovery of adversarial examples. We further introduced innovative techniques and analyses such as Transformer-XL, Autoformalization, the Memorizing Transformer, Batch Size Scaling, and μTransfer. We have worked on and led the development of some of the largest breakthroughs in the field including AlphaStar, AlphaCode, Inception, Minerva, GPT-3.5, and GPT-4.
-
Bard is getting better at logic and reasoning
I believe tuning hyper parameters well without a lot of waste for the largest models was only figured out by Greg Yang/Microsoft Research around 2022 (cited in GPT-4 paper):
https://arxiv.org/abs/2203.03466
Also part of how they predicted the loss ahead of time so well.
-
Cerebras Open Sources Seven GPT models and Introduces New Scaling Law
This is the first time I have seen muP applied by the third party. See Cerebras Model Zoo, where muP models have scale-invariant constant LR.
-
OpenAI’s policies hinder reproducible research on language models
I guess, but its actually not simple to do that, in my experience. There’s another paper on that: https://arxiv.org/abs/2203.03466
Why isn’t chinchilla running google AI chat or whatever then?
-
[D] Anyone else witnessing a panic inside NLP orgs of big tech companies?
Well, but it isn't like this kind of research is new. Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer (2022) tuned hyperparameters in 40M model, transferred it to 6.7B model, and beat OpenAI's 6.7B run. It is likely what OpenAI did is perfecting this kind of research. I note that four authors of that paper (Igor Babuschkin, Szymon Sidor, David Farhi, Jakub Pachocki) are credited for pretraining optimization & architecture at https://openai.com/contributions/gpt-4.
-
[R] Greg Yang's work on a rigorous mathematical theory for neural networks
Tensor Programs I: Wide Feedforward or Recurrent Neural Networks of Any Architecture are Gaussian Processes: https://arxiv.org/abs/1910.12478 Tensor Programs II: Neural Tangent Kernel for Any Architecture: https://arxiv.org/abs/2006.14548 Tensor Programs III: Neural Matrix Laws: https://arxiv.org/abs/2009.10685 Tensor Programs IV: Feature Learning in Infinite-Width Neural Networks: https://proceedings.mlr.press/v139/yang21c.html Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer: https://arxiv.org/abs/2203.03466
- [D] How does one choose a learning rate schedule for models that take days or weeks to train?
- How to do meaningful work as an independent researcher? [Discussion]
-
DeepMind’s New Language Model,Chinchilla(70B Parameters),Which Outperforms GPT-3
I think there remains an immense amount of such suboptimality still hanging from the tree, so to speak.
For example, our recent paper "Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer"[1] shows that even learning rate and initialization used by existing models are deeply wrong. By just picking them correctly (which involves some really beautiful mathematics), we can effectively double the model size of the GPT-3 6.7B model (to be comparable in quality to the 13B model across the suite of benchmark tasks).
Large neural networks behave in a way we are only beginning to understand well just because each empirical probe of any such model is so much more expensive and time consuming than typical models. But principled theory here can have a lot of leverage by pointing out the right direction to look, as it did in our work.
[1] http://arxiv.org/abs/2203.03466
-
"Training Compute-Optimal Large Language Models", Hoffmann et al 2022 {DeepMind} (current LLMs are significantly undertrained)
On the hyperparameter front there seems to be some overlap with the recent hyperparameter transfer paper, which I get the impression Microsoft is going to try to scale, and which was referenced (and so is known) by the authors of this DeepMind paper. Which is to say, there's a good chance we'll be seeing models of this size trained with more optimal hyperparameters pretty soon.
gpt-3
-
GPT4.5 or GPT5 being tested on LMSYS?
>I wasn't talking about "state of the art LLMs," I am aware that commercial offerings are much better trained in Spanish. This was a thought experiment based on comments from people testing GPT-3.5 with Swahili.
A thought experiment from other people comments on another language. So...No. Fabricating failure modes from their constructed ideas about how LLMs work seems to be a frustratingly common occurrence in these kinds of discussions.
>Frustratingly, just few months ago I read a paper describing how LLMs excessively rely on English-language representations of ideas, but now I can't find it.
Most LLMs are trained on English overwhelmingly. GPT-3 had a 92.6% English dataset. https://github.com/openai/gpt-3/blob/master/dataset_statisti...
That the models are as proficient as they are is evidence enough of knowledge transfer clearly happening. https://arxiv.org/abs/2108.13349. If you trained a model on the Catalan tokens GPT-3 was trained on alone, you'd just get a GPT-2 level gibberish model at best.
anyway. These are some interesting papers
How do languages influence each other? Studying cross-lingual data sharing during LLM fine-tuning - https://arxiv.org/pdf/2305.13286
Teaching Llama a New Language Through Cross-Lingual Knowledge Transfer - https://arxiv.org/abs/2404.04042
Multilingual LLMs are Better Cross-lingual In-context Learners with Alignment - https://arxiv.org/abs/2305.05940
It's not like there is perfect transfer but the idea that there's none at all seemed so ridiculous to me (and why i asked the first question). Models would be utterly useless in multilingual settings if that were really the case.
-
What are LLMs? An intro into AI, models, tokens, parameters, weights, quantization and more
Large models: Everything above 10B of parameters. This is where Llama 3, Llama 2, Mistral 8x22B, GPT 3, and most likely GPT 4 sit.
-
Can ChatGPT improve my L2 grammar?
Are generative AI models useful for learning a language, and if so which languages? Over 90% of ChatGPT's training data was in English. The remaining 10% of data was split unevenly between 100+ languages. This suggests that the quality of the outputs will vary from language to language.
-
GPT4 Can’t Ace MIT
I have doubts it was extensively trained on German data. Who knows about GPT4, but GPT3 is ~92% of English and ~1.5% of German, which means it saw more "die, motherfucker, die" than on "die Mutter".
(https://github.com/openai/gpt-3/blob/master/dataset_statisti...)
- Necesito ayuda.
-
[R] PaLM 2 Technical Report
Catalan was 0.018 % of GPT-3's training corpus. https://github.com/openai/gpt-3/blob/master/dataset_statistics/languages_by_word_count.csv.
- I'm seriously concerned that if I lost ChatGPT-4 I would be handicapped
- The responses I got from bard after asking why 100 times… he was pissed 😂
-
BharatGPT: India's Own ChatGPT
>Certainly it is pleasing that they are not just doing Hindi, but some of these languages must be represented online by a very small corpus of text indeed. I wonder how effectively an LLM can be trained on such a small training set for any given language?
as long as it's not the main language it doesn't really matter. Besides English(92.6%), the biggest language by representation (word count) is taken up by french at 1.8%. Most of the languages GPT-3 knows are sitting at <0.2% representation.
https://github.com/openai/gpt-3/blob/master/dataset_statisti...
Competence in the main language will bleed into the rest.
- GPT-4 gets a B on Scott Aaronson's quantum computing final exam
What are some alternatives?
com.openai.unity - A Non-Official OpenAI Rest Client for Unity (UPM)
dalle-mini - DALL·E Mini - Generate images from a text prompt
NTK4A - Code for the paper: "Tensor Programs II: Neural Tangent Kernel for Any Architecture"
DALL-E - PyTorch package for the discrete VAE used for DALL·E.
GP4A - Code for NeurIPS 2019 paper: "Tensor Programs I: Wide Feedforward or Recurrent Neural Networks of Any Architecture are Gaussian Processes"
DALLE-mtf - Open-AI's DALL-E for large scale training in mesh-tensorflow.
cdx-index-client - A command-line tool for using CommonCrawl Index API at http://index.commoncrawl.org/
stylegan2-pytorch - Simplest working implementation of Stylegan2, state of the art generative adversarial network, in Pytorch. Enabling everyone to experience disentanglement
nn - 🧑🏫 60 Implementations/tutorials of deep learning papers with side-by-side notes 📝; including transformers (original, xl, switch, feedback, vit, ...), optimizers (adam, adabelief, sophia, ...), gans(cyclegan, stylegan2, ...), 🎮 reinforcement learning (ppo, dqn), capsnet, distillation, ... 🧠
v-diffusion-pytorch - v objective diffusion inference code for PyTorch.
efficientnet - Implementation of EfficientNet model. Keras and TensorFlow Keras.
dalle-2-preview