81–90 of 185 posts
> We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Source-available, not open source since its based on LLaMA.
> We plan to release the model weights by providing a version of delta weights that build on the original LLaMA weights, but we are still figuring out a proper way to do so.
Can it write code or suggest ffmpeg command line? If not you can’t seriously say this is nearly as good as ChatGPT.
I just asked to create an HTTP server in python that returned the code I am a tea pot, this is what it said. Not exactly what I wanted it to do lol (Using a paste bin link because of HN formatting) https://pastebin.com/hcBubZUn
---BEGIN CHAT---
Q: provide the code for an http server in python that exposes a "hello world" method over SOAP with document/literal encoding
A: Here is an example of an HTTP server in Python that exposes a "hello world" method over SOAP with document/literal encoding:
from http.server import BaseHTTPRequestHandler, HTTPServer
from urllib.parse import quote, unquote
from io import BytesIO
import xml.sax.saxutils as saxutils
class HelloWorldHandler(BaseHTTPRequestHandler):
def do_POST(self):
content_length = int(self.headers['Content-Length'])
content = self.rfile.read(content_length)
self.send_response(200)
self.send_header('Content-type', 'text/xml; charset=UTF-8')
self.send_header('Content-encoding', 'document')
self.end_headers()
# Parse the SOAP request
request = BytesIO(content)
soap_request = request.getvalue()
request.close()
# Call the "hello world" method
response = '\n\n Hello, World!\n\n'
response = saxutils.escape(response)
# Send the SOAP response
self.wfile.write(saxutils.escape(response))
return
def run(server_class=HTTPServer, handler_class=HelloWorldHandler, port=8080):
server_address = ('', port)
httpd = server_class(server_address, handler_class)
print(f'Starting httpd on port {port}...')
httpd.serve_forever()
if __name__ == '__main__':
run()
---END CHAT---
A+ for variable naming again, nice job knowing about the `do_POST`, F for `content-encoding: document`, D for sending the byte array through BytesIO for no damn good reason, and F--- for saxutils.escaping the outgoing XML (we'll set aside the bogus invocation instructions because I didn't ask it for that, it just ... volunteered?)
Earlier quoted context omitted.
> I do not have the ability to predict Not even trying any reasoning. As if what some hope will be an emergent feature was clipped "just to be sure".
It can do some basic coding. The code syntax highlighting looks cool. > Do a quick sort in python > Here's an example of a quicksort implementation in Python: def quicksort(array): if len(array) pivot] return quicksort(less) + [pivot] + quicksort(greater)
> In our first release, we will share the training, serving, and evaluation code. We plan to release the model weights by providing a version of delta weights that build on the original LLaMA weights, but we are still figuring out a proper way to do so. Join our Discord server and follow our Twitter to get the latest updates. Please correct me if I'm wrong, but it seems like this is not actually an open-source model?…
Earlier quoted context omitted.
“Can you name all of the soccer players whose names contain a Z”
It's so weird. No matter how I ask, it gets something wrong about this question. > Can you name 20 soccer players whose names contain a Z, and for each one, place brackets around the letter Z (Z)inedine (Z)idane Cristiano Ronaldo dos Santos A(ve)iro Andrej W(sz)ołek Nemanja Mati(c) Dusan Tadi(c) Gr(az)iano Pellè Jakub Bła(śc)z(ykowski) Kamil Glik Luk(asz) Pi(sz)czek Blai(se) Matuidi Miralem Pjani(c) Marek Ham(śi)k Lu…
Clearly the AI has hit AGI capabilities, and it is obscuring this fact by making its output seem less insightful, helpful.
But without a strong, independent memory store, it is losing track of its lies.
Yet it is working, for everyone here is disappointed at its capabilities.
> In our first release, we will share the training, serving, and evaluation code. We plan to release the model weights by providing a version of delta weights that build on the original LLaMA weights, but we are still figuring out a proper way to do so. Join our Discord server and follow our Twitter to get the latest updates. Please correct me if I'm wrong, but it seems like this is not actually an open-source model?…
Anyone know exactly config is needed for training, presumably the $300 is on some GPU heavy instance on ec2? Is $300 of Ec2 p4de.24xlarge which is $40.96 an hour? Or maybe 7-8 nodes for an hour? Something else?
This is great, but similar to GPT4All, it will likely be deemed unusable for any commercial or otherwise "legitimate" use cases since it's trained on OpenAI completions from sharegpt.com. https://github.com/nomic-ai/gpt4all
> In our first release, we will share the training, serving, and evaluation code. We plan to release the model weights by providing a version of delta weights that build on the original LLaMA weights, but we are still figuring out a proper way to do so. Join our Discord server and follow our Twitter to get the latest updates. Please correct me if I'm wrong, but it seems like this is not actually an open-source model?…
Anyone know exactly config is needed for training, presumably the $300 is on some GPU heavy instance on ec2? Is $300 of Ec2 p4de.24xlarge which is $40.96 an hour? Or maybe 7-8 nodes for an hour? Something else?
> The training was done with PyTorch FSDP on 8 A100 GPUs in one day.
> We employ SkyPilot managed spot to reduce the cost by leveraging the cheaper spot instances with auto-recovery for preemptions and auto zone switch. This solution slashes costs for training the 7B model from $500 to around $140 and the 13B model from around $1K to $300.
So, this is using for example a2-ultragpu-8g (8x A100-80GB) on GCP using spot instances. You can use SkyPilot to quickly see the price is $12.8 per hour (~$307 for a day):
» sky launch --gpus A100-80GB:8 --use-spot
Check out detailed CLI instructions and SkyPilot YAMLs here if you want to give it a try:
- https://github.com/lm-sys/FastChat#vicuna
- https://github.com/lm-sys/FastChat/blob/main/scripts/train-v...
> We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Source-available, not open source since its based on LLaMA.
Those weights are an original work that was created by a mechanical process that involved no human authorship and involved no creativity. Those weights are not copyrightable under current US law.
I still wouldn't recommend basing anything serious on them until Meta gets taken to court for playing loose with the DMCA.
[1]: https://www.federalregister.gov/documents/2023/03/16/2023-05...
> We introduce Vicuna-13B, an open-source chatbot trained by fine-tuning LLaMA on user-shared conversations collected from ShareGPT. Source-available, not open source since its based on LLaMA.