lxe

47 ms·

541.
by lxe 4y ago
Thank you for your amazing work! This is inspired by your whisper wasm demo! I can barely keep up with the changes to your repos. You landed gpt-neox / stablelm support right around the time I was finishing up this demo, which I can’t
542.
by lxe 4y ago
Keep in mind that this demo doesn't have context memory to keep it performant. Every new question starts without the knowledge of previous answers.
543.
by lxe 4y ago
That's pretty impressive. What model are you using?
544.
by lxe 4y ago
I've just added sampling parameters and repetition penalty. Making things a bit more coherent.
545.
by lxe 4y ago
Impressive. This is basically AGI! User: What is 3+3? Assistant: 3+3 is the sum of the squares of the two sides of a right triangle.
546.
by lxe 4y ago
I'm attempting to port repetition penalty to the sampling code to make it less repetitive for gpt-2 based models. Model suggestions welcome! I will be able to get one of the smaller llamas loaded in here.
547.
by lxe 4y ago
I'm using https://huggingface.co/lxe/Cerebras-GPT-1.3B-Alpaca-SP which is https://huggingface.co/cerebras/Cerebras-GPT-1.3B finetuned using LoRA on Alpaca dataset. The cerebras models (gpt-2
548.
by lxe 4y ago
Interesting, I'm getting 100ms/token on plain old wasm with 4 threads via ggml, using a 1.7B quantized cerebras model.
549.
by lxe 4y ago
It is not. It's much slower. But it doesn't require you to use a canary browser with a command line flag, so that's that. EDIT: I guess I should give credit here... 100-200ms/token with a 1.7B model is not slow. Would lo
550.
by lxe 4y ago
Glad to hear it! I added a launch button so it doesn't just download it when you navigate to it
551.
Show HN: WasmGPT – “ChatGPT” in the browser, no WebGPU and no server needed(lxe.co)
185 points by lxe 4y ago|50 comments
552.
by lxe 4y ago
Shameless plug: I built a toy tool to really quickly get into llm finetuning by just pasting your training samples into a textbox: https://github.com/lxe/simple-llm-finetuner
553.
by lxe 4y ago
This certainly led me on a Wikipedia rabbithole into extinct arthopods. My favorite used to be Anomalocaris (anomalous shrimp), until I discovered https://en.wikipedia.org/wiki/Innovatiocaris , the "Innovation Crab
554.
by lxe 4y ago
The model is very good. Looks like it's using GPT-NeoX structure, and the finetuned 7b version is very quick, and beats everything else in the category (vicuna, oasst, openchatkit, gpt4all, etc...) in my "write a poem about cheese
555.
by lxe 4y ago
Finetuning and training models of this size from scratch is now relatively inexpensive and accessible to anyone, especially with all these datasets becoming freely available. This is just the beginning of smaller LLMs that anyone can finetu
556.
by lxe 4y ago
That infinite game of life... just wow!
557.
by lxe 4y ago
How are the models hosted? Edit: they are using their own hosted APIs such as Anthropic or Replicate.
558.
by lxe 4y ago
Haha yup, welcome to the club. I'm getting too old and too distracted.
559.
by lxe 4y ago
Good. With "Xfinity 10G Network," the ISPs are really stretching that truth in advertisement law.
560.
by lxe 4y ago
There's also "jslinux" from Fabrice Bellard himself from a while ago which reused parts of QEMU: https://bellard.org/jslinux/
561.
by lxe 4y ago
Why Vicuna over something like oasst-pythia? Would love to see a table comparing all the new models side by side.
562.
by lxe 4y ago
What a useful and nuanced critique. Thanks!
563.
by lxe 4y ago
This is a great writeup on some of the nuances and gotchas you have to watch out for when finetuning using dreambooth and the generative creative process in general.
564.
by lxe 4y ago
Ah darn it, and here I was trying to do this with ggml and onxx.
565.
by lxe 4y ago
I've been using AutoGPT, which is the other ChatGPT automation tool, and frankly I'm a bit disappointed. It spends most of its time figuring out why system commands are failing and why javascript is blocked, and/or why seleni
566.
by lxe 4y ago
It's a little funny how Microsoft DeepSpeed doesn't fully work on Windows
567.
by lxe 4y ago
Can vouch for vast.ai as well. At these prices anyone can get into llm finetuning at their leisure.
568.
by lxe 4y ago
Why is carrying a firearm such a "hill people choose to die on" item?
569.
by lxe 4y ago
Kudos for the authors for providing the code https://github.com/huawei-noah/KD-NLP/tree/main/DyLoRA and the roberta example. Considering the current state of the OSS LLM community, I'm guessing some
570.
by lxe 4y ago
I'm guessing langchain / llamaindex + openai API?

More ›