A humorous factor occurred after I was enjoying with Poe, a chatbot aggregator from Quora. I chosen the just lately launched free almost-open-source Llama 2 70B Chat mannequin from Meta and gave it the immediate “Generate a Python program to scrape an internet site. Embrace exams for python.org and infoworld.com.” That’s a shorter model of a immediate I used to check CodeWhisperer, Bard, and Copilot X in June. None of these three fashions generated a working program.
To my utter shock, Llama 2 70B Chat aced this take a look at:
IDGThe bot went on to recommend some follow-up questions:
Inform me extra. Are you able to clarify how the BeautifulSoup library works for parsing HTML content material? Are there every other libraries or instruments generally used for net scraping in Python? What are some greatest practices for net scraping to make sure compliance with web site phrases of service?
The generated code regarded good to me, so I copied it into Visible Studio Code, saved it, and ran it. It ran completely:
% /usr/native/bin/python3 /Customers/martinheller/Code/Llama/scrapesite.py martinheller@Martins-M1-MBP Code % /usr/native/bin/python3 /Customers/martinheller/Code/Llama/scrapesite.py Title: Welcome to Python.org Hyperlink: #content material Hyperlink: #python-network Hyperlink: / Hyperlink: https://www.python.org/psf/ Hyperlink: https://docs.python.org Hyperlink: https://pypi.org/ Hyperlink: /jobs/ Hyperlink: /community-landing/ Hyperlink: #prime Hyperlink: / Hyperlink: https://psfmember.org/civicrm/contribute/transact?reset=1&id=2 Hyperlink: #site-map Hyperlink: # Hyperlink: javascript:; Hyperlink: javascript:; Hyperlink: javascript:; Hyperlink: # Hyperlink: https://www.fb.com/pythonlang?fref=ts Hyperlink: https://twitter.com/ThePSF Hyperlink: /neighborhood/irc/ Hyperlink: /about/ Hyperlink: /about/apps/ Hyperlink: /about/quotes/ Hyperlink: /about/gettingstarted/ Hyperlink: /about/assist/ Hyperlink: http://brochure.getpython.information/ Hyperlink: /downloads/ Hyperlink: /downloads/ Hyperlink: /downloads/supply/ Hyperlink: /downloads/home windows/ Hyperlink: /downloads/macos/ Hyperlink: /obtain/different/ Hyperlink: https://docs.python.org/3/license.html Hyperlink: /obtain/alternate options Hyperlink: /doc/ …
Evaluating the Llama-generated code with the CodeWhisperer-generated code, the most important distinction is that Llama used the html.parser mannequin for Stunning Soup, which labored, whereas CodeWhisperer used the lxml mannequin, which choked.
Llama 2 code clarification
I additionally requested Llama 2 70B Chat to elucidate the identical pattern program I had given to CodeWhisperer, Bard, and Copilot X. CodeWhisperer doesn’t at present have a chat window, so it doesn’t do code explanations, however Bard did a terrific job on this process and Copilot X did a superb job.
IDG
IDG
IDGLlama’s clarification (proven above) is nearly as good, or probably higher, than what Bard generated. I don’t utterly perceive why Llama stopped in merchandise 12, however I think that it could have hit a token restrict, except I by chance hit the “cease” button in Poe and didn’t discover.
For extra about Llama 2 typically, together with dialogue of its potential copyright violations and whether or not it’s open supply or not, see “What’s Llama 2? Meta’s massive language mannequin defined.”
Coding with Code Llama
A few days after I completed working with Llama 2, Meta AI launched a number of Code Llama fashions. Just a few days after that, at Google Cloud Subsequent 2023, Google introduced that they have been internet hosting Code Llama fashions (amongst many others) within the new Vertex AI Mannequin Backyard. Moreover, Perplexity made one of many Code Llama fashions out there on-line, together with three sizes of Llama 2 Chat.
So there have been a number of methods to run Code Llama on the time I used to be writing this text. It’s seemingly that there can be a number of extra, and several other code editor integrations, within the subsequent months.
Poe didn’t host any Code Llama fashions after I first tried it, however through the course of writing this text Quora added Code Llama 7B, 13B, and 34B to Poe’s repertoire. Sadly, all three fashions gave me the dreaded “Unable to achieve Poe” error message, which I interpret to imply that the mannequin’s endpoint is busy or not but linked. The next day, Poe up to date, and operating the Code Llama 34B mannequin labored:
IDGAs you may see from the screenshot, Code Llama 34B went one higher than Llama 2 and generated applications utilizing each Stunning Soup and Scrapy.
Perplexity is web site that hosts a Code Llama mannequin, in addition to a number of different generative AI fashions from varied firms. I attempted the Code Llama 34B Instruct mannequin, optimized for multi-turn code technology, on the Python code-generation process for web site scraping:
IDGSo far as it went, this wasn’t a foul response. I do know that the requests.get() technique and bs4 with the html.parser engine work for the 2 websites I urged for exams, and discovering all of the hyperlinks and printing their HREF tags is an effective begin on processing. A really fast code inspection urged one thing apparent was lacking, nonetheless:
IDGNow this seems extra like a command-line utility, however completely different performance is now lacking. I might have most popular a practical kind, however I stated “program” quite than “operate” after I made the request, so I’ll give the mannequin a go. However, this system because it stands will report undefined features when compiled.
IDGReturning JSON wasn’t actually what I had in thoughts, however for the needs of testing the mannequin I’ve in all probability gone far sufficient.
Llama 2 and Code Llama on Google Cloud
At Google Cloud Subsequent 2023, Google Cloud introduced that new additions to Google Cloud Vertex AI’s Mannequin Backyard embody Llama 2 and Code Llama from Meta, and revealed a Colab Enterprise pocket book that allows you to deploy pre-trained Code Llama fashions with vLLM with the most effective out there serving throughput.
If it’s essential to use a Llama 2 or Code Llama mannequin for lower than a day, you are able to do so free of charge, and even run it on a GPU. Use Colab. If you understand how, it’s straightforward. For those who don’t, seek for “run code llama on colab” and also you’ll see a full web page of explanations, together with numerous YouTube movies and weblog posts on the topic. Observe that whereas Colab is free however time-limited and resource-limited, Colab Enterprise prices cash however isn’t restricted.
If you wish to create an internet site for operating LLMs, you need to use the identical vLLM library as used within the Google Cloud Colab Pocket book to arrange an API. Ideally, you’ll set it up on a server with a GPU sufficiently big to carry the mannequin you need to use, however that isn’t completely obligatory: You may get by with one thing like a M1 or M2 Macintosh so long as it has sufficient RAM to run your mannequin. You may also use LangChain for this, at the price of writing or copying a couple of strains of code.
Working Llama 2 with Ollama
In case you are utilizing an Arm-based Macintosh as your workstation, you may run Llama fashions regionally as a command-line utility. The invaluable Sharon Machlis explains how to make use of Ollama; it’s straightforward, though should you don’t have sufficient RAM reminiscence for the mannequin it’ll use digital reminiscence (i.e. SSD or, heaven forfend, spinning disk) and run actually sluggish. (Linux and Home windows assist is deliberate for Ollama.)
I attempted out Ollama with a number of fashions (of the various it helps) on my M1 MacBook Professional, which sadly has solely 8GB of RAM. I began with my normal Python web-scraping code technology process utilizing Llama 2, apparently one of many smaller fashions (7B?). The result’s comparable to what I bought from the Llama 2 70B mannequin operating on Poe, though not as well-structured. Observe that Ollama solely downloads the mannequin the primary time it wants it.
martinheller@Martins-M1-MBP ~ % ollama run llama2
pulling manifest
pulling 8daa9615cce3... 100% |███████████████| (3.8/3.8 GB, 51 MB/s)
pulling 8c17c2ebb0ea... 100% |██████████████████| (7.0/7.0 kB, 39 MB/s)
pulling 7c23fb36d801... 100% |█████████████████| (4.8/4.8 kB, 5.0 MB/s)
pulling 578a2e81f706... 100% |████████████████████| (95/95 B, 129 kB/s)
pulling e35ab70a78c7... 100% |████████████████████| (90/90 B, 550 kB/s)
pulling 96862bb35d77... 100% |██████████████████| (447/447 B, 2.8 MB/s)
verifying sha256 digest
writing manifest
success
>>> Generate a Python program to scrape an internet site. Embrace exams for python.org and infoworld.com
Certain! Right here is an instance of how you need to use the `requests` and `BeautifulSoup` libraries in Python to scrape an internet site:
```
import requests
from bs4 import BeautifulSoup
# Instance web site to scrape
url = "https://www.python.org"
# Ship a request to the web site and get the HTML response
response = requests.get(url)
# Parse the HTML content material utilizing Stunning Soup
soup = BeautifulSoup(response.content material, 'html.parser')
# Discover the precise data you need to scrape on the web site
title = soup.discover('h1', class_='main-title').textual content
print(title) # Output: "Python.org"
# Discover all of the hyperlinks on the web page
hyperlinks = soup.find_all('a')
for hyperlink in hyperlinks:
print(hyperlink.get('href')) # Output: ["/python", "/docs", "/learn", etc.]
# Check the code with completely different web sites
url2 = "https://www.infoworld.com"
response2 = requests.get(url2)
soup2 = BeautifulSoup(response2.content material, 'html.parser')
title2 = soup2.discover('h1', class_='main-title').textual content
print(title2) # Output: "Infoworld"
# Discover all of the hyperlinks on the web page
links2 = soup2.find_all('a')
for hyperlink in links2:
print(hyperlink.get('href')) # Output: ["/news", "/features", "/opinions", etc.]
```
This code sends a request to the desired web site, parses the HTML content material utilizing Stunning Soup, after which finds particular data on the web page (on this case, the title and all of the hyperlinks). You'll be able to modify the `url` variable on the prime of the script to specify the web site you need to scrape.
For instance, if you wish to scrape the Python documentation, you may set `url = "https://www.python.org/doc/` after which use the identical code as earlier than to search out the title and hyperlinks on the web page.
To check the code with completely different web sites, you may modify the `url2` variable on the prime of the script to specify a distinct web site. For instance, you may set `url2 = "https://www.infoworld.com"` to scrape that web site, or every other web site you need to take a look at.
Observe: Be sure you have the mandatory permissions and authorized rights to scrape an internet site. Internet scraping could be unlawful in some instances, so you should definitely examine the phrases of service for the web site you are scraping and respect their guidelines.
With that baseline established, I attempted the identical immediate utilizing Code Llama. Once more, I didn’t specify the mannequin dimension, however it seems like it’s 7B.
martinheller@Martins-M1-MBP ~ % ollama run codellama
pulling manifest
pulling 8268242df6f0... 100% |████████████████████████████████████████████| (3.8/3.8 GB, 50 MB/s)
pulling 2c8743bdc4ad... 100% |███████████████████████████████████████████████| (7.0/7.0 kB, 40 MB/s)
pulling 38fa20ee7daa... 100% |███████████████████████████████████████████████| (4.8/4.8 kB, 28 MB/s)
pulling 578a2e81f706... 100% |█████████████████████████████████████████████████| (95/95 B, 993 kB/s)
pulling 404e21afdc6a... 100% |█████████████████████████████████████████████████| (30/30 B, 187 kB/s)
pulling 86b68a7bb45d... 100% |███████████████████████████████████████████████| (507/507 B, 4.0 MB/s)
verifying sha256 digest
writing manifest
success
>>> Generate a Python program to scrape an internet site. Embrace exams for python.org and infoworld.com
To jot down a Python program that scrapes an internet site, you'll need to make use of a library akin to BeautifulSoup or Scrapy. These libraries let you navigate and parse the HTML construction of a webpage, and extract particular information from it.
Right here is an instance of the way you would possibly use BeautifulSoup to scrape the Python web site at python.org:
```
import requests
from bs4 import BeautifulSoup
# Ship a request to the Python web site
url="https://www.python.org"
response = requests.get(url)
# Parse the HTML content material utilizing Stunning Soup
soup = BeautifulSoup(response.content material, 'html.parser')
# Discover all hyperlinks on the web page
hyperlinks = soup.find_all('a')
# Print every hyperlink to the console
for hyperlink in hyperlinks:
print(hyperlink.get('href'))
```
To jot down exams for this program, you need to use a testing framework akin to unittest or pytest. These frameworks let you outline take a look at features that may be run routinely to confirm that your code is working accurately.
Right here is an instance of the way you would possibly write exams utilizing the unittest module:
```
import unittest
from bs4 import BeautifulSoup
class TestPythonOrg(unittest.TestCase):
def setUp(self):
# Ship a request to the Python web site
url="https://www.python.org"
response = requests.get(url)
# Parse the HTML content material utilizing Stunning Soup
soup = BeautifulSoup(response.content material, 'html.parser')
self.soup = soup
def test_title(self):
title = self.soup.discover('title').textual content
self.assertEqual(title, 'Python Documentation - Python v3.10')
def test_description(self):
description = self.soup.discover('meta', {'identify': 'description'})['content']
self.assertIn('The official homepage for the Python Programming Language', description)
def tearDown(self):
go
if __name__ == '__main__':
unittest.most important()
```
This take a look at program sends a request to the Python web site, parses the HTML content material utilizing Stunning Soup, after which runs two separate exams: one which verifies that the web page title is appropriate, and one other that verifies that the web page description comprises sure key phrases.
You'll be able to run this take a look at program by saving it to a file (e.g. `test_python_org.py`), opening a terminal or command immediate, after which operating the next command:
```
python test_python_org.py
```
It will run the entire take a look at features in this system, and print a message indicating whether or not every take a look at handed or failed.
