Connect with us

SEO

8 Experiments To Improve Organic Traffic

Published

on

8 Experiments To Improve Organic Traffic

Imagine the results you could achieve if you knew the Google search algorithm.

Just think of the organic traffic you could drive to your site if you knew exactly what Google was taking into account when generating search engine results, the precise amount each factor plays on rankings, and precisely what it would take to get to the top. 

But of course, Google would never let you see behind the curtain. Not only could a bad actor use this information for nefarious purposes, but it would take all the fun out of search engine optimization. 

So instead, every time there’s a new algorithm update, it’s up to us to figure out what exactly has changed and how to best leverage this to our advantage. 

And because we in the search engine optimization world are a community, we tend to figure these things out collectively. 

And because there are no (or very few, anyway) absolute answers, a lot of optimization comes down to best practices, theories, and outright guesswork. 

Luckily, there’s a great way to determine if these theories are based on reality or just sheer conjecture. Of course, we’re talking about experimentation.

(Insert mad scientist laugh here.)

Before you get carried away and rush to put on your safety goggles and lab coats, relax. No bubbling beakers full of mystery liquids are involved, and your risk of accidentally creating a monster is very low.

(I would say zero, but I’ve seen the “Terminator” series, and it never explicitly says that a search engine optimizer didn’t create Skynet, so let’s play it safe.)

All your SEO experiments can be done from your desk chair’s comfort (and safety). 

But before we dive into those, let’s first talk about how to test your SEO.

Steps For A Successful SEO Experiment

Thankfully, we don’t have to develop an entirely new framework for devising, conducting, and measuring our SEO tests – we can repurpose the scientific method you’re probably already familiar with. 

Been a while since high school chemistry? Don’t worry; these five steps are sure to sound familiar:

  1. Make an observation. (E.g., my site is not on the first page of Google search results.)
  2. Ask a question and form a hypothesis. Is my meta description optimized for ranking? If I write better meta descriptions, it will improve my ranking.
  3. Gather data. Upon changing these descriptions, track the change in ranking position and site visits.
  4. Analyze the data. Create tables, graphs, and diagrams to help you understand the link between what was changed and changing results.
  5. Draw conclusions. Does the evidence support your prediction? Why or why not? Was site traffic increased due to your new meta descriptions, or did you also receive national news coverage during your testing period?)

You’ll want to do what marketers call A/B testing for accurate results. This means creating two versions of the same page, with only one difference, so you can see which gets the better response. 

Before you start, keep one thing in mind: Incremental changes are essential. If you go wild and change all of these at once, you’ll have no idea which one(s) is making a difference. 

Play it slowly and be patient. Test one, then wait for the results before moving on to the next. This will give you an obvious idea of what you need to do moving forward to ensure you’re always claiming front-page real estate.

With that out of the way, here’s a look at eight different things you can test to improve the ranking of your site:

1. SEO Title

Have you ever written what you thought was an amazing page title, only for Google to rewrite it in search results? That usually happens when the search engine doesn’t feel like your title was a good reflection of the page’s content. 

But you can change this anytime you like.

And even if Google does replace the one you created, Google’s John Mueller confirmed the original title is still used for ranking purposes. This means even if you’re getting your SEO titles changed, it’s still a good idea to try to optimize them.

Here are a few things you can test to see if they generate results:

  • Include your target keyword.
  • Change their length (shorter is not always better).
  • Experiment with brand name positioning or remove it altogether.
  • Get click-baity (e.g., Do you want to lose 20 lbs. fast?).
  • Add published date to demonstrate information relevancy.
  • Get creative (people love what’s new and whimsical).

2. SEO Meta Description

Now, wait just a minute – you’re probably saying right now – Search Engine Journal has been clear that Google hasn’t used meta description in its rankings since sometime between 1999 and 2004. 

Put down your pitchforks. Just because they aren’t a direct factor in SERPs doesn’t mean meta descriptions aren’t an essential SEO element. 

For example, they can help improve your click-through rate, compel searchers, generate brand exposure, and help differentiate you from the competition. And all of these impact user behavior, which is a signal that Google factors in. 

There have been entire articles written about creating awesome meta descriptions. But for our purposes, here are a few things you can A/B test to see if you can improve your organic traffic:

  • Try different lengths. Traditional SEO wisdom suggests character count between 156-165 – see what works for you.
  • Add keywords.
  • Change your tone. Generally speaking, your style should match your brand’s voice, but for a specific page, maybe this isn’t the case.
  • Get specific – are you getting a lot of visits from one long-tail keyword? Add that to your meta description. 

3. Internal Anchor Text

You may know anchor text as the visible, clickable (usually blue) text in a hyperlink like this. 

Not only is this useful for giving additional context to users, but Google has confirmed: Anchor text helps it better understand a page’s content, allowing it to rank those pages for relevant searches. 

For example, in the paragraph above, “Google has confirmed” is the anchor text attached to the target link, which directs to a link proving that confirmation. 

There are several types of anchor text you can use, including those with exact or partial-match keywords, branded (Search Engine Journal), images, generic (“click here”), and naked links (https://www.searchenginejournal.com).

To experiment with the impact, internal anchor text can have on your organic traffic. You can try things like:

  • Changing their length (though shorter is often better).
  • Adding keywords, particularly low-density keywords.
  • Increasing specificity.
  • Changing them to be more target link-specific.

4. Schema Markup

Schema markup is a type of structured data used by Google and other search engines.

Following guidelines established by Schema.org, it is essentially a lingua franca for search engines. It’s an established standard that uses a unique vocabulary to help search engines more clearly understand your content. 

It’s used to create rich snippets for adding information about events, recipes, people, videos, reviews, and products, among other things. These, in turn, make your link appear more prominent in SERPs. 

And while schema is not directly factored into your ranking score, like SEO title, it can improve your click-through rate and impact your user behavior scores. 

Experiment with adding schema markup to your pages and see if it improves your results.

5. Images

A picture is worth a thousand words – everyone knows that. This is because humans are visual creatures. And web designers and SEO professionals have recognized the importance of including images on webpages for a long time. 

And it’s not just because they add visual interest and grab attention; they can also improve your search ranking.

Original (not stock), high-quality images optimized for SEO can reap the rewards. 

Here are some things you can try with your images to improve your traffic:

  • Add images. You should add images to every page. No one wants to read a wall of text.
  • Choose a different file name. You want to immediately make it clear to Google what your image is depicting. Try adding your keywords.
  • Change your formatting. There is no one-size-fits-all for digital images. You may want to change your file type depending on your image needs. JPEGs are good for larger photos. PNG files preserve background transparency. 
  • Compress files when possible. In general, smaller files are always better.
  • Ensure responsiveness. With mobile search being such an essential factor in modern SEO, you want to ensure your images always look great on phones and desktops. 
  • Add alt text if the image can’t be displayed for some reason. 

6. Headers

Your headlines and subheads give your page structure. This makes it easier for humans to browse and for search engines to understand what each section is about. 

Google’s John Mueller was very clear about this, emphasizing the presence of any headings, not just H1s, sends a strong signal about the page’s content. It would be best if you took special care with all your H tags, from H1 down to H6. This is because they also serve as an accessibility aid and navigational tool in addition to their structural benefits. 

Once upon a time, your headers were a massive factor in your ranking. But then, like always, people abused them, and Google started cracking down on keyword stuffing, overuse, and other dirty heading tricks. 

That said, while they are far from the only factor Google takes into account, there is no question they are essential.

Here are a few things you can play with to try and improve your SEO results:

  • Add more headings (except H1s). Improve the structure and clarity of your content by adding more subheads.
  • Try using them to break up blocks of text. No one likes massive text blocks. H2s, H3s, etc., are the perfect solution.
  • Add keywords. Like nearly everything else we’ve discussed, you should experiment with the use of keywords in your headings. 
  • Optimize for featured snippets. Grab attention and draw new traffic by claiming those special boxes on search results. Write your headers to land these.
  • Get creative. Try making your headings and subheads more interesting. 

7. Word Count

 You already know content is the essential thing in any SEO strategy. But you may not have thought much about how the number of words you use can impact your ranking and traffic.

While you shouldn’t expect word count to push you over the top and take you from page six to the top result, it can help define your site as relevant and valuable to a search query – and draw in readers. 

What do we mean by this? Again, it’s not a direct ranking factor, according to John Mueller.

With that said, longer-form copy tends to rank higher. This is because using more words provides Google with more information on what your page is about. 

And if these longer pieces are well-written (like the one you’re reading in the author’s ever-so-humble opinion), they’ll help establish you as an authority on the topic. 

And experimentation is easy. Take one of your existing articles or blog posts and duplicate it. On the second one, expound at greater length upon your topic. 

Please note, we’re not talking about using your word count tricks from college (e.g., “at a later date” instead of just “later”). Instead, you should expand on ideas and topics, add examples and cite additional sources.

Then, see which one performs better on search engines. Chances are, it will be the longer one. 

Read this piece for more on using word count for SEO optimization. 

8. URL

Some SEO experts will swear URLs with keywords perform better than generic ones. Is this the case?

Well, yes and no.

On the one hand, Google has confirmed when it performs its initial crawl of a site, keywords in URLs help it understand what the site is about.

However, as this is only factored in when a new site is crawled, its role in an ongoing SEO strategy is minimal. 

But again, this doesn’t mean you can’t use them to your advantage. Clear URLs create a better user experience and can be used as naked anchor tags much more accessible than one with 75 random numbers and letters tacked on at the end. 

Experiment with your URLs. Take some of your old, non-descriptive links and add keywords to them. Shorten long URLs.

If you’re worried about losing links from the old page, add a 301-redirect pointing to the new one. 

Don’t Be Afraid To Try Something New

Search engine optimization is a constantly shifting landscape. Changing trends can change how people interact with your website.

As algorithms change and new technologies emerge, your strategy needs to evolve. 

There has never been, and probably never will be, “set it and forget it” search engine optimization. It will always require forward-thinkers and people willing to experiment to find new ways to get their websites to the top of the rankings. 

Who says you can’t be one of them? Someone had to be the first to figure out that keyword stuffing helped rankings, and someone else had to figure out when it stopped working.

If you’re willing to experiment and try new things, you may find the next brilliant new strategy. Just don’t forget to share it with us.

More Resources:


Featured Image: Tatyana Vyc/Shutterstock



Source link

Keep an eye on what we are doing
Be the first to get latest updates and exclusive content straight to your email inbox.
We promise not to spam you. You can unsubscribe at any time.
Invalid email address

SEO

Technical SEO Checklist for 2024: A Comprehensive Guide

Published

on

Technical SEO Checklist 2024 Comprehensive Strategies

Technical SEO Checklist 2024 Comprehensive Strategies

With Google getting a whopping total of six algorithmic updates and four core updates in 2023, you can bet the search landscape is more complicated (and competitive) to navigate nowadays.

To succeed in SEO this year, you will need to figure out what items to check and optimize to ensure your website stays visible. And if your goal is to not just make your website searchable, but have it rank at the top of search engine results, this technical SEO checklist for 2024 is essential.

Webmaster’s Note: This is part one of our three-part SEO checklist for 2024. I also have a longer guide on advanced technical SEO, which covers best practices and how to troubleshoot and solve common technical issues with your websites.

Technical SEO Essentials for 2024

Technical SEO refers to optimizations that are primarily focused on helping search engines access, crawl, interpret, and index your website without any issues. It lays the foundation for your site to be properly understood and served up by search engines to users.

1. Website Speed Optimization

A site’s loading speed is a significant ranking factor for search engines like Google, which prioritize user experience. Faster websites generally provide a more pleasant user experience, leading to increased engagement and improved conversion rates.

Server Optimization

Often, the reason why your website is loading slowly is because of the server it’s hosted on. It’s important to choose a high-quality server that ensures quick loading times from the get-go so you skip the headache that is server optimization.

Google recommends keeping your server response time under 200ms. To check your server’s response time, you need to know your website’s IP address. Once you have that, use your command prompt.

In the window that appears, type ping, followed by your website’s IP address. Press enter and the window should show how long it took your server to respond. 

If you find that your server goes above the recommended 200ms loading time, here’s what you need to check:

  1. Collect the data from your server and identify what is causing your response time to increase. 
  2. Based on what is causing the problem, you will need to implement server-side optimizations. This guide on how to reduce initial server response times can help you here.
  3. Measure your server response times after optimization to use as a benchmark. 
  4. Monitor any regressions after optimization.

If you work with a hosting service, then you should contact them when you need to improve server response times. A good hosting provider should have the right infrastructure, network connections, server hardware, and support services to accommodate these optimizations. They may also offer hosting options if your website needs more server resources to run smoothly.

Website Optimization

Aside from your server, there are a few other reasons that your website might be loading slowly. 

Here are some practices you can do:

  1. Compressing images to decrease file sizes without sacrificing quality
  2. Minimizing the code, eliminating unnecessary spaces, comments, and indentation.
  3. Using caching to store some data locally in a user’s browser to allow for quicker loading on subsequent visits.
  4. Implementing Content Delivery Networks (CDNs) to distribute the load, speeding up access for users situated far from the server.
  5. Lazy load your web pages to prioritize loading the objects or resources only your users need.

A common tool to evaluate your website speed is Google’s PageSpeed Insights or Google Lighthouse. Both tools can analyze the content of your website and then generate suggestions to improve its overall loading speed, all for free. There are also some third-party tools, like GTMetrix, that you could use as well.

Here’s an example of one of our website’s speeds before optimization. It’s one of the worst I’ve seen, and it was affecting our SEO.

slow site speed score from GTMetrixslow site speed score from GTMetrix

So we followed our technical SEO checklist. After working on the images, removing render-blocking page elements, and minifying code, the score greatly improved — and we saw near-immediate improvements in our page rankings. 

site speed optimization results from GTMetrixsite speed optimization results from GTMetrix

That said, playing around with your server settings, coding, and other parts of your website’s backend can mess it up if you don’t know what you’re doing. I suggest backing up all your files and your database before you start working on your website speed for that reason. 

2. Mobile-First Indexing

Mobile-first Indexing is a method used by Google that primarily uses the mobile version of the content for indexing and ranking. 

It’s no secret that Google places a priority on the mobile users’ experience, what with mobile-first indexing being used. Beyond that, optimizing your website for mobile just makes sense, given that a majority of people now use their phones to search online.

This change signifies that a fundamental shift in your approach to your website development and design is needed, and it should also be part of your technical SEO checklist.

  1. Ensuring the mobile version of your site contains the same high-quality, rich content as the desktop version.
  2. Make sure metadata is present on both versions of your site.
  3. Verify that structured data is present on both versions of your site.

Tools like Google’s mobile-friendly test can help you measure how effectively your mobile site is performing compared to your desktop versions, and to other websites as well.

3. Crawlability & Indexing Check

Always remember that crawlability and Indexing are the cornerstones of SEO. Crawlability refers to a search engine’s ability to access and crawl through a website’s content. Indexing is how search engines organize information after a crawl and before presenting results.

  1. Utilizing a well-structured robots.txt file to communicate with web crawlers about which of your pages should not be processed or scanned.
  2. Using XML sitemaps to guide search engines through your site’s content and ensure that all valuable content is found and indexed. There are several CMS plugins you can use to generate your sitemap.
  3. Ensuring that your website has a logical structure with a clear hierarchy, helps both users and bots navigate to your most important pages easily. 

Google Search Console is the tool you need to use to ensure your pages are crawled and indexed by Google. It also provides reports that identify any problems that prevent crawlers from indexing your pages. 

4. Structured Data Markup

Structured Data Markup is a coding language that communicates website information in a more organized and richer format to search engines. This plays a strategic role in the way search engines interpret and display your content, enabling enhanced search results through “rich snippets” such as stars for reviews, prices for products, or images for recipes.

Doing this allows search engines to understand and display extra information directly in the search results from it.

Key Takeaway

With all the algorithm changes made in 2023, websites need to stay adaptable and strategic to stay at the top of the search results page. Luckily for you, this technical SEO checklist for 2024 can help you do just that. Use this as a guide to site speed optimization, indexing, and ensuring the best experience for mobile and desktop users.

Source link

Keep an eye on what we are doing
Be the first to get latest updates and exclusive content straight to your email inbox.
We promise not to spam you. You can unsubscribe at any time.
Invalid email address
Continue Reading

SEO

Why Google Seems To Favor Big Brands & Low-Quality Content

Published

on

By

Why Google Seems To Favor Big Brands & Low-Quality Content

Many people are convinced that Google shows a preference for big brands and ranking low quality content, something that many feel has become progressively worse. This may not be a matter of perception, something is going on, nearly everyone has an anecdote of poor quality search results. The possible reasons for it are actually quite surprising.

Google Has Shown Favoritism In The Past

This isn’t the first time that Google’s search engine results pages (SERPs) have shown a bias that favored big brand websites. During the early years of Google’s algorithm it was obvious that sites with a lot of PageRank ranked for virtually anything they wanted.

For example, I remember a web design company that built a lot of websites, creating a network of backlinks, raising their PageRank to a remarkable level normally seen only in big corporate sites like IBM. As a consequence they ranked for the two-word keyword phrase, Web Design and virtually every other variant like Web Design + [any state in the USA].

Everyone knew that websites with a PageRank of 10, the highest level shown on Google’s toolbar, practically had a free pass in the SERPs, resulting in big brand sites outranking more relevant webpages. It didn’t go unnoticed when Google eventually adjusted their algorithm to fix this issue.

The point of this anecdote is to point out an instance of where Google’s algorithm unintentionally created a bias that favored big brands.

Here are are other  algorithm biases that publishers exploited:

  • Top 10 posts
  • Longtail “how-to” articles
  • Misspellings
  • Free Widgets in footer that contained links (always free to universities!)

Big Brands And Low Quality Content

There are two things that have been a constant for all of Google’s history:

  • Low quality content
  • Big brands crowding out small independent publishers

Anyone that’s ever searched for a recipe knows that the more general the recipe the lower the quality of recipe that gets ranked. Search for something like cream of chicken soup and the main ingredient for nearly every recipe is two cans of chicken soup.

A search for Authentic Mexican Tacos results in recipes with these ingredients:

  • Soy sauce
  • Ground beef
  • “Cooked chicken”
  • Taco shells (from the store!)
  • Beer

Not all recipe SERPs are bad. But some of the more general recipes Google ranks are so basic that a hobo can cook them on a hotplate.

Robin Donovan (Instagram), a cookbook author and online recipe blogger observed:

“I think the problem with google search rankings for recipes these days (post HCU) are much bigger than them being too simple.

The biggest problem is that you get a bunch of Reddit threads or sites with untested user-generated recipes, or scraper sites that are stealing recipes from hardworking bloggers.

In other words, content that is anything but “helpful” if what you want is a tested and well written recipe that you can use to make something delicious.”

Explanations For Why Google’s SERPs Are Broken

It’s hard not to get away from the perception that Google’s rankings for a variety of topics always seem to default to big brand websites and low quality webpages.

Small sites grow to become big brands that dominate the SERPs, it happens. But that’s the thing, even when a small site gets big, it’s now another big brand dominating the SERPs.

Typical explanations for poor SERPs:

  • It’s a conspiracy to increase ad clicks
  • Content itself these days are low quality across the board
  • Google doesn’t have anything else to rank
  • It’s the fault of SEOs
  • Affiliates
  • Poor SERPs is Google’s scheme to drive more ad clicks
  • Google promotes big brands because [insert your conspiracy]

So what’s going on?

People Love Big Brands & Garbage Content

The recent Google anti-trust lawsuit exposed the importance of the Navboost algorithm signals as a major ranking factor. Navboost is an algorithm that interprets user engagement signals to understand what topics a webpage is relevant for, among other things.

The idea of using engagement signals as an indicator of what users expect to see makes sense. After all, Google is user-centric and who better to decide what’s best for users than the users themselves, right?

Well, consider that arguably the the biggest and most important song of 1991, Smells Like Teen Spirt by Nirvana, didn’t make the Billboard top 100 for that year. Michael Bolton and Rod Stewart made the list twice, with Rod Stewart top ranked for a song called “The Motown Song” (anyone remember that one?)

Nirvana didn’t make the charts until the next year…

My opinion, given that we know that user interactions are a strong ranking signal, is that Google’s search rankings follow a similar pattern related to users’ biases.

People tend to choose what they know. It’s called a Familiarity Bias.

Consumers have a habit of choosing things that are familiar over those that are unfamiliar. This preference shows up in product choices that prefer brands, for example.

Behavioral scientist, Jason Hreha, defines Familiarity Bias like this:

“The familiarity bias is a phenomenon in which people tend to prefer familiar options over unfamiliar ones, even when the unfamiliar options may be better. This bias is often explained in terms of cognitive ease, which is the feeling of fluency or ease that people experience when they are processing familiar information. When people encounter familiar options, they are more likely to experience cognitive ease, which can make those options seem more appealing.”

Except for certain queries (like those related to health), I don’t think Google makes an editorial decision to certain kinds of websites, like brands.

Google uses many signals for ranking. But Google is strongly user focused.

I believe it’s possible that strong user preferences can carry a more substantial weight than Reviews System signals. How else to explain why Google seemingly has a bias for big brand websites with fake reviews rank better than honest independent review sites?

It’s not like Google’s algorithms haven’t created poor search results in the past.

  • Google’s Panda algorithm was designed to get rid of a bias for cookie cutter content.
  • The Reviews System is a patch to fix Google’s bias for content that’s about reviews but aren’t necessarily reviews.

If Google has systems for catching low quality sites that their core algorithm would otherwise rank, why do big brands and poor quality content still rank?

I believe the answer is that is what users prefer to see those sites, as indicated by user interaction signals.

The big question to ask is whether Google will continue to rank what users biases and inexperience trigger user satisfaction signals.  Or will Google continue serving the sugar-frosted bon-bons that users crave?

Should Google make the choice to rank quality content at the risk that users find it too hard to understand?

Or should publishers give up and focus on creating for the lowest common denominator like the biggest popstars do?



Source link

Keep an eye on what we are doing
Be the first to get latest updates and exclusive content straight to your email inbox.
We promise not to spam you. You can unsubscribe at any time.
Invalid email address
Continue Reading

SEO

Google Announces Gemma: Laptop-Friendly Open Source AI

Published

on

By

Google Announces Gemma: Laptop-Friendly Open Source AI

Google released an open source large language model based on the technology used to create Gemini that is powerful yet lightweight, optimized to be used in environments with limited resources like on a laptop or cloud infrastructure.

Gemma can be used to create a chatbot, content generation tool and pretty much anything else that a language model can do. This is the tool that SEOs have been waiting for.

It is released in two versions, one with two billion parameters (2B) and another one with seven billion parameters (7B). The number of parameters indicates the model’s complexity and potential capability. Models with more parameters can achieve a better understanding of language and generate more sophisticated responses, but they also require more resources to train and run.

The purpose of releasing Gemma is to democratize access to state of the art Artificial Intelligence that is trained to be safe and responsible out of the box, with a toolkit to further optimize it for safety.

Gemma By DeepMind

The model is developed to be lightweight and efficient which makes it ideal for getting it into the hands of more end users.

Google’s official announcement noted the following key points:

  • “We’re releasing model weights in two sizes: Gemma 2B and Gemma 7B. Each size is released with pre-trained and instruction-tuned variants.
  • A new Responsible Generative AI Toolkit provides guidance and essential tools for creating safer AI applications with Gemma.
  • We’re providing toolchains for inference and supervised fine-tuning (SFT) across all major frameworks: JAX, PyTorch, and TensorFlow through native Keras 3.0.
  • Ready-to-use Colab and Kaggle notebooks, alongside integration with popular tools such as Hugging Face, MaxText, NVIDIA NeMo and TensorRT-LLM, make it easy to get started with Gemma.
  • Pre-trained and instruction-tuned Gemma models can run on your laptop, workstation, or Google Cloud with easy deployment on Vertex AI and Google Kubernetes Engine (GKE).
  • Optimization across multiple AI hardware platforms ensures industry-leading performance, including NVIDIA GPUs and Google Cloud TPUs.
  • Terms of use permit responsible commercial usage and distribution for all organizations, regardless of size.”

Analysis Of Gemma

According to an analysis by an Awni Hannun, a machine learning research scientist at Apple, Gemma is optimized to be highly efficient in a way that makes it suitable for use in low-resource environments.

Hannun observed that Gemma has a vocabulary of 250,000 (250k) tokens versus 32k for comparable models. The importance of that is that Gemma can recognize and process a wider variety of words, allowing it to handle tasks with complex language. His analysis suggests that this extensive vocabulary enhances the model’s versatility across different types of content. He also believes that it may help with math, code and other modalities.

It was also noted that the “embedding weights” are massive (750 million). The embedding weights are a reference to the parameters that help in mapping words to representations of their meanings and relationships.

An important feature he called out is that the embedding weights, which encode detailed information about word meanings and relationships, are used not just in processing input part but also in generating the model’s output. This sharing improves the efficiency of the model by allowing it to better leverage its understanding of language when producing text.

For end users, this means more accurate, relevant, and contextually appropriate responses (content) from the model, which improves its use in conetent generation as well as for chatbots and translations.

He tweeted:

“The vocab is massive compared to other open source models: 250K vs 32k for Mistral 7B

Maybe helps a lot with math / code / other modalities with a heavy tail of symbols.

Also the embedding weights are big (~750M params), so they get shared with the output head.”

In a follow-up tweet he also noted an optimization in training that translates into potentially more accurate and refined model responses, as it enables the model to learn and adapt more effectively during the training phase.

He tweeted:

“The RMS norm weight has a unit offset.

Instead of “x * weight” they do “x * (1 + weight)”.

I assume this is a training optimization. Usually the weight is initialized to 1 but likely they initialize close to 0. Similar to every other parameter.”

He followed up that there are more optimizations in data and training but that those two factors are what especially stood out.

Designed To Be Safe And Responsible

An important key feature is that it is designed from the ground up to be safe which makes it ideal for deploying for use. Training data was filtered to remove personal and sensitive information. Google also used reinforcement learning from human feedback (RLHF) to train the model for responsible behavior.

It was further debugged with manual re-teaming, automated testing and checked for capabilities for unwanted and dangerous activities.

Google also released a toolkit for helping end-users further improve safety:

“We’re also releasing a new Responsible Generative AI Toolkit together with Gemma to help developers and researchers prioritize building safe and responsible AI applications. The toolkit includes:

  • Safety classification: We provide a novel methodology for building robust safety classifiers with minimal examples.
  • Debugging: A model debugging tool helps you investigate Gemma’s behavior and address potential issues.
  • Guidance: You can access best practices for model builders based on Google’s experience in developing and deploying large language models.”

Read Google’s official announcement:

Gemma: Introducing new state-of-the-art open models

Featured Image by Shutterstock/Photo For Everything



Source link

Keep an eye on what we are doing
Be the first to get latest updates and exclusive content straight to your email inbox.
We promise not to spam you. You can unsubscribe at any time.
Invalid email address
Continue Reading

Trending

Follow by Email
RSS