A large-scale illegal gold mining area in the buffer zone of the Tambopata National Reserve in the Madre de Dios region of Peru, captured by the CBERS-4 satellite in September 2019. The image has been rendered in false color to highlight the mining pits: forest appears black, yellow marks exposed pits, and water-filled ponds show a range of colors depending on their sediment content. This area is the same region used in the change-analysis example in Ai2's blog, but this image is not OlmoEarth output. Image: Coordenação-Geral de Observação da Terra (Brazil's National Institute for Space Research, INPE), via Wikimedia Commons, licensed under CC BY-SA 2.0.

The Allen Institute for AI (Ai2) released the satellite embedding model OlmoEarth v1.3 on October 22, 2026, alongside global annual embedding data at 10-meter resolution and a browser-based exploration tool that requires no login. Ai2 says its 128-dimensional embeddings scored highest in 20 of 27 comparisons against AlphaEarth Foundations and TESSERA v2. As of October 23, no independent third-party evaluation of v1.3 could be found, so these results remain Ai2's own assessment.

The Model and Accompanying Data

OlmoEarth v1.3 is an Earth observation embedding model: it reads a full year of radar (Sentinel-1) and multispectral (Sentinel-2, Landsat) satellite time series for a given location and outputs a 128-dimensional vector for each 10-meter pixel, the first 64 dimensions of which can also be used on their own as a smaller embedding. An embedding can be thought of as a “digital summary” of a patch of land: by training a small classifier on top of it, users can produce crop maps, land cover maps, or change analyses without downloading imagery themselves or renting GPUs to fine-tune a model. Previously, OlmoEarth was mainly used through task-specific fine-tuning in OlmoEarth Studio; v1.3 is the first version to make embeddings its primary deliverable.

The Hugging Face model card (allenai/OlmoEarth-v1_3-Base) states that the backbone is a ViT-Base with 1.28 hundred million (one hundred twenty-eight million) parameters, plus a Perceiver with 1430 ten-thousand (fourteen point three million) parameters and a linear “student” layer with about 10 ten-thousand (one hundred thousand) parameters, the latter compressing 768-dimensional features down to 128 dimensions; the global annual embeddings are generated by this model stack. The technical report adds that v1.3's backbone architecture is identical to v1.2 Base and that its training data is also carried over from v1.2. What is new are two components: an attention pooler that aggregates multi-temporal, multi-sensor features into a single vector, and a distillation layer that compresses 768 dimensions down to 128 and 64 dimensions.

The precomputed embeddings released with the model are hosted on Source Cooperative and can be read directly with any Zarr client, without an API key. According to the dataset description, as of the initial release on October 22, 2023—2025 has global coverage, while 2017—2022 covers only some countries; Ai2 says the rest will be filled in “in the coming weeks,” and the exact completion date remains to be verified. According to the technical report, coverage spans land and some nearshore waters from 59.5°S to 83.6°N, excluding Antarctica, with a single year of data amounting to about 150 TB; each 20 km tile also comes with a provenance record that traces back to the original satellite scenes used. The OlmoEarth Embeddings Explorer, launched at the same time, requires no registration and lets users switch between years, find similar locations, train a classifier with a small number of labeled points, and export GeoTIFFs in the browser.

License and Use Restrictions

The model weights, embedding data, and code are all released under the OlmoEarth Artifact License, the same as v1 and v1.2. The license field in the Hugging Face metadata is marked “other,” and the repository includes the full text in LICENSE.txt. The license permits free use, modification, distribution, and the creation of derivative models, but Section 2 prohibits two categories of use: first, military and defense-related uses, with the text specifically naming weapons development, military operations, intelligence gathering, and surveillance and policing of people; second, extractive activities, including planning or assisting the extraction of oil, natural gas, and minerals, with drilling, mining, and deforestation given as examples. Distributors must credit Ai2 as the source and include or link to the license, and derivative works must also carry these two restrictions forward.

Thus, the “fully open” described in the blog means that the weights, pretraining data, and pretraining code are all available and reproducible; it does not mean there are no use restrictions. The August 2026 survey “Earth Embeddings” likewise places OlmoEarth in the category of models that use a custom license restricting certain application scenarios. The blog uses the embeddings to track the expansion and decline of illegal gold mining in Peru, yet the license does not spell out, case by case, where projects such as monitoring or enforcement support fall relative to the prohibited uses above. Whether a specific project is restricted requires checking against the original license text or confirming with Ai2; this article cannot determine that, and it remains to be verified.

Official Benchmark Results and How Comparisons Were Counted

Ai2 compares against two models: AlphaEarth Foundations (AEF, 64-dimensional), released by Google DeepMind in 2025, and the University of Cambridge team's TESSERA v2 (the report uses TESSERA-v2-Large, 128-dimensional). The blog states that across 27 evaluations spanning six continents, the 128-dimensional embeddings scored highest on 20 and ranked first overall, while the 64-dimensional version led on 18.

The main text of the technical report uses a different count: looking only at the kNN-20 classifier, the 128-dimensional embeddings were best on 10 of 11 tasks, and the 64-dimensional embeddings beat both baselines on 7. These 11 tasks comprise the 8 balanced-sampling datasets from the AEF paper, plus 3 tests on two crop segmentation datasets, PASTIS and PLANTEUR. The exception is an Ethiopian crop dataset, on which the report says TESSERA v2 consistently outperforms OlmoEarth v1.3. No summary table in the report could be found showing how the blog's “27 evaluations” break down; this remains to be verified.

Data source: Table 2 of the Ai2 technical report; evaluations of all three embeddings were run by Ai2 itself.

mIoU of OlmoEarth v1.3 (128-dim and 64-dim), TESSERA v2, and AEF on three crop segmentation tests. Drawn with code (Chart.js configuration, rendered via QuickChart), with data taken from Table 2 of the OlmoEarth v1.3 technical report; all are Ai2's own evaluations, not independent results.

There are several things to keep in mind when reading these numbers. First, on PASTIS the 128-dimensional version is only 0.3 points higher than TESSERA v2. Second, PLANTEUR is a new benchmark the Ai2 team introduced alongside this release, with labels from five regions in France's overseas departments; the “leave-one-island-out” test trains the classifier on the remaining islands and tests on the held-out island, to measure performance when moving to an unfamiliar region. Third, the three embeddings differ in size, with AEF having only 64 dimensions. Fourth, the AEF embeddings were taken from Google Earth Engine and TESSERA v2 inference was run following its official repository's instructions; both are reproductions by Ai2. The report also notes that the embeddings used for evaluation are not bit-for-bit identical to the publicly released ones; their mIoU on the PASTIS validation set is 59.8 and 60.4 respectively, a small difference.

On cost, Table 5 of the report shows that pretraining v1.3 Base took 539 H100 GPU hours, about 53% of v1.2 Base (1012 hours) and 18% of v1 Base (2989 hours), with carbon emissions of about 0.07 tonnes of carbon dioxide equivalent (which the report describes as a lower-bound estimate). The blog estimates that reproducing the final round of pretraining would cost about $2720 in H100 rental fees. However, computing one year of global embeddings takes about 1.4 ten-thousand (roughly fourteen thousand) H100 hours, about 26 times the pretraining cost. On fine-tuning, the blog says v1.3 performs on par with v1.2; the report gives an average rank across all tasks of 4.34, slightly behind v1.2 Base's 4.02 and ahead of v1 Base's 4.74.

The State of Third-Party Evaluation

As of October 23, 2026, this article could not find any independent reproduction or third-party benchmark of OlmoEarth v1.3. The Hugging Face model card does not include a table of evaluation results. Coverage by TPS Report and Unite.AI both relay Ai2's numbers; TPS Report states explicitly that these results come from Ai2's own evaluations and lists independent reproduction as a key open question, while the Unite.AI article notes that it was AI-generated.

The external comparisons that can be found all concern older versions. The University of Cambridge TESSERA team's July 2026 TESSERA v2 paper included OlmoEarth v1 Large (3.08 hundred million, i.e., three hundred eight million, parameters) in its comparison: across its 14 held-out datasets, TESSERA v2 had an overall score of 0.647, OlmoEarth-L and TESSERA v1 were tied at 0.614, and AEF scored 0.590; when switching to a linear probe that reads only the center-point features, however, OlmoEarth-L held a slight lead on the CityRep urban task. This evaluation comes from a competing team and did not test v1.3; it only shows that with a different team, task set, or readout method, the rankings may change.

The field has already been discussing these issues. The August 2026 survey “Earth Embeddings” cites an audit finding that the reported accuracy of the same model on the same benchmark can differ by more than ten percentage points; the authors therefore advocate for an independently maintained, unified evaluation framework and leaderboard, and caution that randomly split training and test sets are insufficient to show that a model transfers to new regions. For readers planning to use these embeddings, the prudent approach is to compare several embeddings and simple baselines on their own target region and task, using geographically partitioned validation sets, before deciding whether to adopt them.

References

1. Ai2, OlmoEarth v1.3: State-of-the-art embeddings for the whole planet, Hugging Face blog, 2026-10-22.

2. Ai2, OlmoEarth v1.3: State-of-the-art embeddings for the whole planet, Ai2 official blog, 2026-10-22.

3. Team OlmoEarth, OlmoEarth v1.3 Embeddings, Allen Institute for AI technical report, 2026.

4. Ai2, allenai/OlmoEarth-v1_3-Base, Hugging Face model card, accessed 2026-10-23.

5. Ai2, OlmoEarth Artifact License, model repository LICENSE.txt, accessed 2026-10-23.

6. OlmoEarth, OlmoEarth v1.3 Embeddings, Source Cooperative dataset description, accessed 2026-10-23.

7. Brown et al., AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data, arXiv:2507.22291, 2025.

8. Feng et al., TESSERA v2: Scaling Pixel-wise Earth Foundation Models, arXiv:2607.03949, 2026.

9. Stewart et al., Earth Embeddings, arXiv:2608.03410, 2026.

10. TPS Report, Ai2 releases OlmoEarth v1.3 satellite embedding model with open global 10m embeddings, 2026-10-22.

11. Unite.AI, Ai2 Releases OlmoEarth V1.3 With Global 10M Satellite Embeddings, 2026-10-22.

12. Smart Academy · Your own academy for the age of intelligence

13. Compiled and published by Scholay, the one-stop academic platform; if there is any infringement, please let us know and it will be taken down.

14. Also published on EnterScholar, a community of scholars for the age of intelligence

AI-generated illustration: the four stages of data, training, deployment, and monitoring, shown alongside a certificate, corresponding to the end-to-end machine learning workflow assessed by the Professional Machine Learning Engineer exam. The image contains no real interfaces, data, or trademarks.

The current exam guide for Google Cloud's Professional Machine Learning Engineer certification is dated June 1, 2026. Google officially states that the exam has been updated to reflect the renaming of Vertex AI to Gemini Enterprise Agent Platform, and generative AI content is spread across the six exam domains. The exam lasts 2 hours and has 50–60 questions. The registration fee is $200, the exam is offered in English and Japanese, and the certification is valid for 2 years. All information below was verified on October 23, 2026.

Target Audience and Exam Format

The official page describes the target candidates for this certification as engineers who use Google Cloud capabilities and traditional machine learning methods to build, evaluate, productionize, and optimize AI solutions. Specific requirements include handling large-scale, complex data, writing reusable code, and designing and operating AI solutions based on foundation models. Candidates should also be familiar with model architectures, data and machine learning pipelines, MLOps, and metric interpretation, and should have a basic understanding of prompt and context engineering, application development, and data governance. The exam has no prerequisites. Google recommends at least 3 years of industry experience, including at least 1 year designing and managing solutions on Google Cloud. The page also notes that the exam does not directly test programming ability; basic proficiency in Python and SQL is enough to understand questions that include code snippets.

Regarding eligible regions, the exam terms list only the countries and regions where the exam cannot be taken and do not separately address whether test centers are available in mainland China. Candidates in mainland China need to check the registration system for their location to find out whether they can book an in-person test center and what network requirements apply to online proctoring. This article lists these points as pending verification for now.

Six Exam Domains and Their Weights

The current exam guide is labeled as the version “as of June 1, 2026.” The certification homepage explains that this update reflects the transition from Vertex AI to Gemini Enterprise Agent Platform and changes in Google Cloud's data and analytics product line, and that it gives priority to Google Cloud–native solutions. Accordingly, the guide uses new names such as Agent Platform AutoML, Agent Platform Pipelines, and Model Registry. If you are studying with older materials that use Vertex AI naming, you need to check product names item by item against the new guide.

The six percentages add up to 101% because the guide gives only approximate figures. By weight, training, deployment, and pipelines (items 3 through 5) together account for about 59% and form the core of the exam. Generative AI is not listed as a separate domain. Instead, it appears across the topics in each domain, for example fine-tuning Gemini in BigQuery, using LLM-as-a-judge to evaluate generative solutions, and using Model Armor to guard against malicious prompts.

Validity, Renewal, and Retakes

The exam terms state that professional-level certifications are valid for 2 years from the date of issue, unless the exam description specifies otherwise. According to the renewal options page, all professional-level certifications except Professional Cloud Architect and Professional Data Engineer, including this one, can currently be renewed only by retaking the standard exam. The renewal window opens 60 days before the certification expires, and passing the exam extends validity by 2 years. Google Skills continuing-education renewal for these certifications is marked “not yet available, coming soon,” and no shorter renewal exam is offered. If you retake the exam before the window opens, your result may be rejected and the exam fee will not be refunded. Within 30 days after a certification expires, you can still renew and keep your original certification number; after more than 30 days, you must earn the certification again at full price. When you first earn the certification, you receive a half-price renewal code, which you can view in your CM Connect profile. Starting October 29, 2026, CM Connect will no longer support email-and-password sign-in, and you will need to sign in with a Google Skills or Google Skills for Partners account instead.

The retake rules are the same for associate-level and professional-level exams: you can take an exam at most 4 times within two years. After a first failed attempt you must wait 14 days, after a second failed attempt you must wait 60 days, and after a third failed attempt you must wait 365 days. You must pay the fee again for every attempt, and these rules also apply to renewal exams. The exam terms further prohibit any use of AI software during the exam, as well as sharing or using leaked exam content.

Differences from Related Certifications and Preparation Resources

Associate Cloud Engineer and Generative AI Leader are often compared with this certification, but the three are designed for different roles. Associate Cloud Engineer is aimed at engineers who deploy, secure, and operate applications and infrastructure in the cloud, and it does not focus on machine learning. Generative AI Leader is open to any role, with or without hands-on technical experience, and focuses on generative AI fundamentals, Google Cloud generative AI products, and business strategy. Professional Machine Learning Engineer, by contrast, requires candidates to make the full set of engineering decisions on Google Cloud, from data and training to deployment and monitoring.

Official preparation resources follow the five steps listed on the certification homepage. First, gain hands-on experience; you can use the Google Cloud free tier to try some products within the monthly limits. Next, assess your skills against the exam guide. Then work through the official sample questions to get familiar with the question format. After that, study the Professional Machine Learning Engineer Certification learning path on Google Skills. The path is maintained by Google Cloud and consists of 17 activities, including on-demand courses, labs, and skill badges. Finally, schedule an online or test-center exam. The homepage also links to webinars. Third-party question banks are not official resources, and the exam terms explicitly prohibit the use of unauthorized exam content circulating online.

References

1. Google Cloud. Professional ML Engineer Certification | Learn | Google Cloud, accessed 2026-10-23.

2. Google Cloud. Professional Machine Learning Engineer Certification exam guide as of June 1, 2026, accessed 2026-10-23.

3. Google Cloud Certification Help. Certification Renewal Options, accessed 2026-10-23.

4. Google Cloud. Google Cloud Certification Exam Policies and Exam Terms and Conditions, accessed 2026-10-23.

5. Google Cloud. Associate Cloud Engineer Certification | Learn | Google Cloud, accessed 2026-10-23.

6. Google Cloud. Generative AI Leader | Learn | Google Cloud, accessed 2026-10-23.

7. Google Skills. Professional Machine Learning Engineer Certification | Google Skills, accessed 2026-10-23.

8. Smart Academy · Your own academy for the age of intelligence

9. This information was compiled and published by Scholay One-Stop Academic Platform. If any content infringes your rights, please let us know and we will take it down.

10. This information is also published on EnterScholar, a community for scholars in the age of intelligence.

Screenshot of the Google Colab web notebook interface. Image: Wikideas1, source Wikimedia Commons, license CC0 1.0 (Public Domain Dedication), uploaded on August 5, 2025. This screenshot was not taken by this site, and the interface may differ from the version current in October 2026.

When quota allows, Google Colab's free tier can use an NVIDIA T4 GPU, but notebooks run for at most 12 hours, idle sessions are reclaimed, and execution stops once you close the browser. The paid Colab Pro grants 100 compute units per month, Pro+ grants 600, and Pro+ supports background execution of up to 24 hours. Paid plans are sold only in the 67 countries and regions officially listed, and mainland China is not among them. The information in this article was verified on October 23, 2026.

GPUs, TPUs and Session Limits on the Free Tier

The official Colab FAQ states that Colab is free to use and provides computing resources including GPUs and TPUs, but these resources are not guaranteed and not unlimited. The official notebook "Making the Most of your Colab Subscription" explains that the free version can use an NVIDIA T4 GPU when quota and availability allow. As for TPUs, the FAQ only says that the types of GPUs and TPUs available change over time; which kind of TPU the free tier can currently be assigned is yet to be verified.

On session length, free notebooks can run for at most 12 hours, and how long they actually run also depends on resource availability and individual usage patterns. A runtime times out after being idle for a while, and virtual machines also have a maximum lifetime. The FAQ explicitly warns that users without a paid subscription should not expect background execution: once you stop interacting with the page, execution is interrupted and the virtual machine is then deleted. You can think of the free tier as a public seat in a library: as long as you are sitting there studying, you can keep using it, but if you step away too long the seat is taken back, and anything you left on the desk is cleared away.

Colab does not publish specific usage caps, idle timeout durations, maximum virtual machine lifetimes or available GPU types, on the grounds that these parameters fluctuate over time. The FAQ also lists two types of restrictions: mining cryptocurrency, downloading torrents, using multiple accounts to get around resource limits, creating deepfakes and similar activities are prohibited on all managed runtimes; and on free runtimes without a positive compute unit balance, remote control such as SSH, bypassing the notebook interface in favor of a web UI, chess training and distributed computing nodes may be terminated at any time without further warning.

Compute Units, Pricing and Regions for Pro and Pro+

Colab offers three self-service paid options: Pay As You Go, Colab Pro and Colab Pro+. The official pricing page lists that Pro grants 100 compute units per month and gives access to faster GPUs and the highest-memory machine types. Pro+ adds another 500 per month on top of that, for a total of 600, offers priority access to high-end GPUs, and supports background execution: as long as the compute unit balance is positive, a running notebook can keep going for up to 24 hours after the browser is closed. Whether Pro users get longer execution times, and whether outputs can be saved to Google Drive, depends on availability. Pay As You Go does not require a subscription.

The rules for compute units are the same across all three options: they expire after 90 days and, except where required by law, are non-refundable and non-transferable. Once the balance runs out, the account falls back to free-tier rules, and running tasks may be terminated by the backend. Officially there is no fixed conversion for how much GPU time one compute unit buys; Google only says it depends on the virtual machine configuration and the GPU or TPU used.

As for pricing, the specific region-dependent amounts shown on the official pricing page could not be read from the page text during this verification, so official prices are yet to be verified. GPU rental provider QuantaCloud, in an article dated September 28, 2026, recorded the following US prices: Pro at 9.99 US dollars per month, Pro+ at 49.99 US dollars per month, and Pay As You Go at 9.99 US dollars for 100 units or 49.99 US dollars for 500 units. This is a record from a third-party commercial website that this site has not rechecked against official pages, and local-currency prices in other regions are likewise yet to be verified. The Additional Terms also state that subscription price increases will be notified by email in advance, and that consumers in the European Economic Area, Brazil, Israel, South Korea and Turkey have a statutory withdrawal right of 7 to 14 days.

Regarding regions, the FAQ lists a total of 67 countries and regions where paid plans are sold, including the United States, the United Kingdom, Japan, India, Singapore, Hong Kong (China) and Taiwan (China); mainland China and South Korea are not on the list. Users outside the listed areas cannot purchase Colab paid plans.

There are two other related routes. The first is Google AI paid plans: the FAQ explains that subscribing to an eligible Google AI plan also grants monthly Colab compute units, which count toward the same balance as units from Colab's own subscriptions; Google One plans that include only storage, as well as trial periods, do not include this benefit. The Google One help page states that Google AI Pro members can receive 200 Colab compute units, available only to users aged 18 and over, and likewise only effective in regions where Colab paid plans are offered. The second is Colab Pro for Education, a free one-year plan for students and faculty at US universities, which is currently no longer accepting new applications.

Comparison of benefits across Colab's free tier, Pay As You Go, Pro and Pro+. AI-generated illustrative chart; the data in the table comes from the Google Colab FAQ and pricing page, verified on 2026-10-23; official prices are yet to be verified.

Key Points of the Data and Privacy Terms

Using Colab requires accepting both the Google Terms of Service and the "Google Colab Additional Terms of Service" (last modified on May 14, 2024), and the Additional Terms also recommend reading the Google Privacy Policy. The Additional Terms prohibit sharing your account with others or using someone else's account, and also prohibit selling, renting or otherwise providing Colab access to third parties. Generative code features are labeled as experimental; users must test the code themselves and comply with open-source license requirements.

Notebooks are saved in Google Drive. When you share a notebook, its text, code, outputs and comments are shared along with it. If you do not want to share outputs, you can check the option to omit code cell output when saving (Omit code cell output when saving this notebook) under Edit > Notebook settings. Custom files and libraries in the virtual machine are not shared with the notebook. Once Google Drive is mounted, any code in the notebook can access all the files in your Google Drive, so before running a notebook shared by someone else, it is best to read through the code once first.

Colab's built-in generative AI features are available only to accounts aged 18 or older. The FAQ explains that when you use these features, Google collects prompts, related code, generated output, feature usage information and feedback in order to provide, improve and develop Google's products and machine learning technologies, including enterprise products such as Google Cloud; human reviewers may read, annotate and process this data. The data is stored in a way that Google cannot identify who provided it, after which deletion requests can no longer be honored, and it is retained for up to 18 months. Google recommends not including confidential information or personally identifiable information in prompts or feedback. By default, Colab AI does not read Google Drive files or user Secrets, but it will generate code that accesses them if you explicitly ask it to.

Practical Tips for Learning Models with Colab

First get your data processing and model code working on a CPU runtime, and switch to a GPU only once you are ready to train. Google notes that running code in a GPU runtime does not necessarily mean you are actually using the GPU; when you are not using the GPU, you should switch back to a standard runtime to avoid hitting GPU usage limits. You can run nvidia-smi to see which graphics card you have been assigned, and when you don't need it, change the accelerator to None under Runtime > Change runtime type. Disconnecting the runtime and closing the tab promptly when you are done also lowers the chance of triggering usage limits.

Design your training workflow on the assumption that it can be cut off at any moment. The free tier has a 12-hour cap and reclaims idle sessions, so training scripts should periodically save checkpoints to Google Drive and resume from the most recent checkpoint after a disconnection. Package datasets as .zip or .tar.gz where possible, copy them to the virtual machine's local disk and extract them there, to avoid frequently reading and writing large numbers of small files in the mounted directory; when the root of your Google Drive holds more than roughly ten thousand items, mounting may fail.

Match practice projects to the free resources. Training small networks and LoRA or QLoRA fine-tuning of small models are better suited to the free tier, and the official example notebooks offer ready-made starting points such as LoRA and QLoRA fine-tuning of Gemma and pretraining a Transformer from scratch with KerasHub. Tasks that require long, unattended training runs are not a good fit for the free tier.

Think through your needs before paying. If you only occasionally need to run a heavier task, Pay As You Go requires no subscription; you only need Pro+ if you want training to continue after closing the browser. If your region is not on the paid-plan list, you can, as the FAQ describes, connect to a local runtime that you control, but this approach cannot mount Google Drive. Do not rotate through multiple accounts to get around limits; this is explicitly prohibited on all managed runtimes.

References

1. Google. Google Colab FAQ. No publication date shown on the page; accessed 2026-10-23.

2. Google. Colab Paid Services Pricing. Accessed 2026-10-23.

3. Google. Google Colab Additional Terms of Service. Last modified 2024-05-14; accessed 2026-10-23.

4. Google Colab. Making the Most of your Colab Subscription. Accessed 2026-10-23.

5. Google One Help Center. Use Google AI Pro benefits. Accessed 2026-10-23.

6. Faiz Ahmed (QuantaCloud). Google Colab Pricing vs Renting a GPU: When to Switch. 2026-09-28; third-party commercial website, used only as a lead for US pricing.

7. Smart Academy · Your own academy for the intelligent era

8. Compiled and published by Scholay, the all-in-one academic platform. If this infringes your rights, please let us know and it will be taken down.

9. Also published on EnterScholar, a community of scholars for the intelligent era

A Google TPU 3.0 device. Kaggle's documentation once listed the TPU v3-8 as its free TPU model; the current model has yet to be verified. Image: Zinskauf, via Wikimedia Commons, licensed under CC BY-SA 4.0.

Kaggle gives registered users free cloud compute in Notebooks. According to the official documentation, the GPU quota is 30 hours per week (sometimes higher, depending on demand and resources), and TPUs are capped at 20 hours per week. A single session can run for up to 12 hours on CPU or GPU and up to 9 hours on TPU, with 20 GB of auto-saved disk space. This article uses Kaggle's official documentation to summarize the quotas, hardware, verification requirements, and terms of use. The information was verified on October 22, 2026.

Weekly Quotas and Available Hardware

Kaggle's GPU usage page says the free GPU quota is calculated and reset weekly. The quota is 30 hours, "sometimes higher, depending on demand and resources." The TPU page lists a maximum of 20 hours per week and up to 9 hours per session. Neither page says what day of the week or what time the quota resets, so the exact reset time has yet to be verified. You can check your remaining quota in the editor's settings menu, at the top of the kaggle.com/notebooks page, on your profile page, and in the session management window. The documentation says both GPUs and TPUs are free but does not list eligible regions, so any regional restrictions have yet to be verified. At busy times, you may have to wait in a queue for either one.

The values in the table come from the "Technical Specifications" section of the Kaggle Notebooks documentation, which notes that they reflect the configuration "at the time of writing."

The official documentation does not list GPU memory size, so this has yet to be verified. The documentation is inconsistent about the TPU model. The "Adding a free TPU" section of the Notebooks documentation names the TPU v3-8, and the TPU page also uses the v3-8 as its example. However, the same TPU page notes that its content was written for "older TPUs that Kaggle no longer supports." The TPU models shown in the editor's Accelerator menu are the ones currently available; the official wording has yet to be verified.

Kaggle also runs a limited-time promotion labeled as experimental. If you link your Kaggle account to an active Colab Pro or Pro+ subscription, you get an extra 15 or 30 hours of GPU time per week, respectively. This time does not use Colab compute units and runs on Kaggle's existing hardware. The promotion terms say that Kaggle may change the quota or end the promotion at any time, and the end date may not be announced publicly or in advance, so the deadline has yet to be verified. This article did not verify the cost of a Colab Pro subscription itself.

Session Length, Disk, and Memory

The technical specifications section of the Notebooks documentation lists the resource limits for each session: CPU and GPU sessions can run for up to 12 hours, and TPU sessions for up to 9 hours. When you save a version with "Save & Run All," the whole notebook must also run from start to finish within 12 hours (9 hours for TPU). If your training takes longer than this, you need to split it into several segments yourself.

Disk space comes in two parts. /kaggle/working provides 20 GB of auto-saved space. A notebook can save up to 20 GB of output, which you can later mount again as input in other notebooks. Outside that directory there is also temporary space, which is not kept after the session ends; the documentation does not state its capacity, so this has yet to be verified. For uploaded datasets, each dataset is capped at 200 GB, and your private datasets are also capped at 200 GB in total.

The documentation gives two different idle timeouts. The technical specifications section says an interactive session ends if you neither edit nor run any code for 20 minutes. The GPU usage tips page, however, says interactive sessions stay active until they reach a 60-minute idle limit. Which figure is current has yet to be verified. To be safe, stop the session manually in the Active Events window at the bottom left of the page before you step away. The GPU usage tips page also notes that doing this can save up to 60 minutes of compute time.

The NVIDIA Tesla P100 (GP100). The P100 is one of the free GPU options listed in Kaggle's technical specifications. Image: FritzchensFritz, via Wikimedia Commons, licensed under CC0 1.0.

Verification Requirements, Internet Access, and Terms of Use

We have not found any official Kaggle rule that requires phone verification specifically for GPUs or TPUs. None of the three documents (Notebooks, GPU, and TPU) says that "you must verify your phone number first." The third-party fast.ai course resources page says that you need a phone-verified account to use GPUs on Kaggle, and that you can do this under "Phone Verification" on the account page. Kaggle's own Benchmarks documentation requires phone verification before an account can use resources such as the LLM API quota. It also says that accounts registered after December 15, 2025 need additional identity verification to run task notebooks, but this applies only to Benchmarks. For the verification requirements for GPUs, TPUs, and internet access, follow the prompts that actually appear on your account settings page; the official rules have yet to be verified. In addition, the Terms of Use allow each person to have only one account, so do not register multiple accounts to get more quota.

Internet access is controlled by the Internet toggle in the editor's Session options. The documentation says internet access is on by default for new notebooks, and when it is on, you can install extra packages with !pip install. Some code competitions require submitted notebooks to have internet access turned off. In that case, you can use the Dependency Manager to package dependencies in advance and install them offline. The official documentation does not say whether the internet toggle also requires phone verification, so this has yet to be verified. Notebooks can also connect to Google Cloud services such as BigQuery and Google Cloud Storage. These services bill the linked Google Cloud account and are not covered by the free quota.

The Terms of Use (effective June 22, 2025) say that the service may be used only for personal, internal, non-commercial purposes, and they explicitly prohibit cryptocurrency mining. The Acceptable Use Policy, effective on the same date, prohibits abusing resources, specifically including cryptocurrency mining, DDoS attacks, server farming, and "activities unrelated to machine learning or data science." Using free compute to host services unrelated to learning over the long term may violate these rules, and violating the terms can lead to account termination.

Datasets and models each have their own licenses. The uploader chooses the license for a public dataset. Common licenses include Creative Commons, GPL, and ODC, and any license not in the drop-down list is stated in the dataset description. The Terms of Use require you to follow the copyright notices on any content you access, and they prohibit copying, republishing, or otherwise using that content without the rights holder's consent. To use competition data, you must first accept that competition's rules, and the competition rules also determine whether you can use external data. Some models (the documentation uses Gemma as an example) require you to agree to a custom license first. Gated models may also ask you to fill in some information, and you can use them only after the access status shows "Accepted."

Tips for Learning AI and Running Experiments

The first four tips below come from Kaggle's official GPU usage tips page. The others are practices based on the session-length and storage limits above, plus official recommendations from the TPU page.

• Turn on the GPU only when you actually need it. The official tips note that libraries such as pandas and scikit-learn do not run faster just because a GPU is available. You can first get your data processing and code working in a CPU session, then switch to a GPU for training.
• Don't use "Save & Run All" just to save your work. It reruns the whole notebook from start to finish. If you only want to keep your progress, use Quick Save or simply download the .ipynb file.
• Clean up sessions promptly. If you submit the same notebook again before the previous submission has finished, you will get multiple batch sessions running in parallel; you can cancel the run that uses the old code. Stop the interactive session before you close the page.
• Push versions with the Kaggle API. This lets you submit a new notebook version to run without opening an interactive session at all.
• Plan training around the session limit. Treat 12 hours (9 hours for TPU) as your limit and regularly write checkpoints to /kaggle/working. In the next session, mount the previous notebook's output as input and resume training from the checkpoint.
• For TPU code, follow the newer documentation. The examples on the TPU page mainly use TensorFlow/Keras. The page recommends organizing training data into tens to hundreds of TFRecord files, each tens to hundreds of MB in size. However, the page notes that it was written for older TPUs, so when you migrate code, follow Google Cloud's official TPU documentation.
• Train on TPU, submit on GPU. Some code competitions do not accept submissions run on TPUs. The official workaround is to train and save the model on a TPU in one notebook and save the output as a dataset. You then load the model in a GPU notebook to run inference and submit.
• Start from public notebooks. You can fork public notebooks directly from competition and dataset pages, reproduce their results first, and then modify them. Any models attached to the original notebook are carried over into your copy.

References

1. Kaggle. Notebooks | How to Use Kaggle. No update date shown on the page; accessed 2026-10-22.

2. Kaggle. Efficient GPU Usage Tips. No update date shown on the page; accessed 2026-10-22.

3. Kaggle. Tensor Processing Units (TPUs). No update date shown on the page; accessed 2026-10-22.

4. Kaggle. Datasets | How to Use Kaggle. Accessed 2026-10-22.

5. Kaggle. Models | How to Use Kaggle. Accessed 2026-10-22.

6. Kaggle. Benchmarks | How to Use Kaggle. Accessed 2026-10-22.

7. Kaggle. Terms of Use. Effective 2025-06-22; accessed 2026-10-22.

8. Kaggle. Acceptable Use Policy. Effective 2025-06-22; accessed 2026-10-22.

9. fast.ai. Practical Deep Learning for Coders - Kaggle. Third-party course page; accessed 2026-10-22.

10. Smart Academy · Your own academy for the age of intelligence

11. Compiled and published by Scholay, the all-in-one academic platform. If any content infringes your rights, please let us know and it will be taken down.

12. Also published on EnterScholar, a community for scholars in the age of intelligence

Speech recognition models process speech picked up by microphones (illustrative photo, not directly related to IBM). Photo: Panos Sakalakis; source: Wikimedia Commons; license: CC BY 2.0.

On October 21, 2026, IBM published an update to its Granite Speech models on the Hugging Face blog. It added granite-speech-5.0-2b and updated two existing models, Turbo CTC and 4.1 2B Plus. All three are released under the Apache 2.0 license. On the Open ASR Leaderboard's English short-form audio benchmark, 5.0 2B has an average word error rate of 5.06%. By the leaderboard data, that places it 8th overall and 2nd among open-weight models, but the ranking mixes averages taken over different numbers of test sets.

What Was Released and What Each Model Is For

IBM's Granite speech team produced the release, and the blog byline includes researchers such as George Saon and Brian Kingsbury. The blog gives each of the three models a use case. 5.0 2B is for high-accuracy multilingual transcription and translation. Turbo CTC is for high-throughput and on-device transcription. 4.1 2B Plus is for multi-speaker recordings such as meetings and phone calls. The dates do not fully agree. The blog is dated October 21, but the 5.0 2B model card says October 20, and the discussion on the leaderboard PR shows the model was still not public on October 17. Which date counts as the official public release has yet to be verified.

The information in the table comes from the three official model cards, and all three models are licensed under Apache 2.0. Note that "Attribute" (speaker attribution) in the blog title refers to a capability of 4.1 2B Plus; the 5.0 2B model card does not list speaker labeling. 5.0 2B uses an intermediate checkpoint of granite-4.0-1b-base as its language model backbone. It was trained on about 390,000 hours of public audio, plus synthetic data created for Japanese recognition, keyword biasing, and translation. The model card recommends cutting long recordings into segments of about 2 minutes or less before transcription.

Results Reported by IBM

All figures in this section come from IBM's blog and model cards and are officially reported results. For 5.0 2B, the blog says its average word error rate (WER, lower is better) on the Open ASR Leaderboard's English short-form public test sets is 5.06%, which ranked 2nd among open-weight models as of October 21, 2026. For punctuation and capitalization, the blog says capitalization F1 ranges from 89.7 to 99.5. However, the model card table lists 5.0 2B's values as 90.16 to 99.51, and 89.71 is actually the score of the previous-generation 4.1 2B. Punctuation error rate went down on 7 of 9 test sets and rose slightly on CommonVoice English and French. Results for noisy and far-field conditions (the FFASR leaderboard), multilingual recognition, and speech translation appear only as charts in the model card. This article could not read the exact values, so they have yet to be verified.

For Turbo CTC, the model card says the new multilingual version reaches an English average WER of 5.51%, better than the English-only version's 5.78%. It also notes that 5.51% is an unofficial result and that the leaderboard still lists the English-only version. The blog says the model reaches 19,870 RTFx on the leaderboard (processing speed as a multiple of the audio's real-time length), about 1.6 times the 12,600 it had at launch. In the leaderboard data read for this article, however, the entry is listed as 21,134.31. The two figures do not match, and the reason has yet to be verified.

For 4.1 2B Plus, the blog compares the old and new versions (all figures are error rates, so lower is better):

The per-test breakdown in the model card shows the gains are not across the board. WDER got worse on FISHER (from 0.9 to 1.0) and CALLHOME English (from 2.2 to 3.0) but improved on AMI-SDM and GALE. On the leaderboard's revised English test sets, average WER rose from 6.01% to 6.10%. The model card does not explain how the blog's summary WDER values of 4.40 and 4.10 were calculated, so this has yet to be verified.

Where It Stands on the Open ASR Leaderboard

The Open ASR Leaderboard is hosted under Hugging Face's hf-audio organization, and its result data are public in the open-asr-leaderboard-results dataset. On October 22, 2026, this article read english_short_latest.csv from that dataset. At the time, the latest commit was f3f9949, the one that added the 5.0 2B results. The English short-form leaderboard currently uses 9 public test sets: cleaned versions of AMI, Earnings22, Gigaspeech, and VoxPopuli; LibriSpeech clean and other; SPGISpeech; URGENT2024; and Voice Arena Monsoon (Indian English). The leaderboard does not state a formal methodology version number, so this has yet to be verified.

This result does not rest on IBM's word alone. On October 14, IBM researcher gsaon opened PR #245 in the leaderboard's code repository and attached run logs from Hugging Face Jobs that showed an average WER of 5.06% across the 9 test sets. On October 19, repository collaborator Deep-unlearning said in a comment that they had run the public test sets independently on an H200 and got the same 5.06%. The PR was merged on October 21 and the result went live. However, the reproduction logs were not attached to the PR, and no results on the leaderboard's private test sets have been published. Both points have yet to be verified.

When the CSV's average WER column is sorted from lowest to highest, 5.0 2B (5.06%) is 8th. Of the 7 entries above it, 6 are closed-source commercial services, and OtterAI speech-v3.1 (3.88%) is 1st. The only open-weight model among them is Microsoft's Phi-4-multimodal-instruct (5.02%, MIT license). Going by this column, the blog's claim of "2nd among open-weight models" matches the leaderboard data.

The ranking does not compare like with like, however. This article recalculated each row from the per-test-set values in the CSV and found that the average column covers different numbers of test sets. For 5.0 2B it covers all 9. Phi-4-multimodal has no URGENT2024 result, so its average covers 8. For older entries such as the previous-generation 4.1 2B and Qwen3-ASR-1.7B-hf, the listed values match the average over 7 test sets, with the two LibriSpeech test sets left out. Averaged over the same 9 test sets, Qwen3-ASR-1.7B-hf scores 4.94%, lower than 5.0 2B's 5.06%. On the 8 test sets where Phi-4-multimodal has results, 5.0 2B averages 4.42%, lower than Phi-4-multimodal's 5.02%. In other words, with a consistent method 5.0 2B is still 2nd among open-weight models, but the model ahead of it is Qwen3-ASR-1.7B-hf. Its 8th place overall also depends on the method. These figures are this article's own calculations, not the leaderboard's official ranking. Why the leaderboard mixes different methods has yet to be verified.

Compared with the previous generation on the same 9 test sets, 4.1 2B has an average WER of 5.35% and 5.0 2B has 5.06%. WER went down on 8 of the 9 test sets; only LibriSpeech other rose slightly, from 2.19% to 2.21% (this article's calculation from the CSV). The two models are about equally fast, with RTFx of 546.84 and 544.52 in the CSV, respectively. IBM reported an RTFx of 570.17 for 5.0 2B in the PR log. Whether the gap from the leaderboard figure comes from different hardware has yet to be verified.

Spectrogram of the English word "open" being spoken (illustrative image, unrelated to the leaderboard data in this article). According to the Granite Speech 5.0 2B model card, its speech encoder takes log-mel spectrogram features as input, which are the same kind of time-frequency representation. Author: Kelly Fitz; source: Wikimedia Commons; license: CC BY-SA 3.0.

References

1. IBM Granite. Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster. Hugging Face Blog, 2026-10-21.

2. IBM Granite. ibm-granite/granite-speech-5.0-2b model card. Hugging Face, accessed 2026-10-22.

3. IBM Granite. ibm-granite/granite-speech-5.0-470m-turboctc model card. Hugging Face, accessed 2026-10-22.

4. IBM Granite. ibm-granite/granite-speech-4.1-2b-plus model card. Hugging Face, accessed 2026-10-22.

5. Hugging Face hf-audio. open-asr-leaderboard-results: english_short_latest.csv. Commit f3f9949, read 2026-10-22.

6. huggingface/open_asr_leaderboard. Add granite-speech-5.0-2b (PR #245). GitHub, merged 2026-10-21.

7. Smart Academy · Your own academy for the age of intelligence

8. Compiled and published by Scholay All-in-One Academic Platform. If this content infringes your rights, please let us know and it will be taken down.

9. Also published on EnterScholar, a community of scholars for the age of intelligence

The Stanford University campus seen from Hoover Tower (photographed May 19, 2014). Photo: Pmlineditor. Source: Wikimedia Commons. License: CC BY-SA 4.0.

The latest offering of Stanford CS336 "Language Modeling from Scratch" is the Spring 2026 term. Its course website, YouTube lecture recordings and GitHub assignments are all public, so you can follow along for free. The five assignments begin with writing a tokenizer and a Transformer, then cover GPU systems optimization, scaling laws, pretraining data processing and alignment, in that order. The barriers to entry are solid Python and PyTorch skills, a large amount of coding time and GPU costs that you pay yourself. Public self-study does not come with a certificate.

Public Materials and Course Dates

The course website, cs336.stanford.edu, is currently labeled "Stanford / Spring 2026". It also keeps archive pages for the two previous spring offerings, 2025 and 2024. The Spring 2026 offering is taught by Percy Liang and Tatsunori Hashimoto. It starts on March 30 and runs for 19 sessions through June 3, and the last two sessions are guest lectures. The schedule attaches lecture materials to each of the first 17 lectures. Percy's lectures come as executable Python lecture notes (lecture_XX.py), and Tatsu's lectures come as PDF slides. The recordings are collected in the YouTube playlist "Stanford CS336: Language Modeling from Scratch | Spring 2026", and the website's schedule also has links next to the lecture titles.

The course code is hosted under the GitHub organization stanford-cs336. The lectures repository holds the lecture source code for this offering, and assignment1-basics through assignment5-alignment are the student versions of the five assignments. Lecture notes from earlier offerings are in two separate repositories, spring2024-lectures and spring2025-lectures. A GitHub check on October 22, 2026 showed that the course website repository's most recent commit was on June 1, 2026. The lectures repository's most recent push was on September 12, 2026, and assignment1-basics's most recent push was on April 7, 2026. The Assignment 1 repository uses the MIT License, but the lectures repository does not declare a license, so check separately before republishing or adapting the lecture materials. There is also one small discrepancy. The README titles for Assignments 2 through 5 all say Spring 2026, but the README title on Assignment 1's main branch still says Spring 2025. Before you start, use the code and preview links for Assignment 1 in the website's schedule as the authoritative version.

The Five Assignments and Their Lectures

The course is built around five assignments, and the lecture sequence largely follows them. The table below is compiled from the website's Spring 2026 schedule. The dates are the release and due dates for that offering, and the website notes that the due dates are tentative.

The schedule also includes Lecture 10 on inference, Lecture 12 on evaluation and the two guest lectures on June 1 and 3 (Daniel Selsam and Dan Fu). These sessions do not map directly to an assignment. Self-learners may also find the website's rules for enrolled students useful. The assignment handouts are self-contained, so you should not look at existing implementations online unless a handout says otherwise. You may ask ChatGPT or similar tools low-level programming or conceptual questions, but you may not have them solve the problems for you. The course also strongly recommends turning off AI autocomplete tools such as Cursor Tab and GitHub Copilot, because they make it hard to understand the material deeply.

Prerequisites, Compute Costs and Certificates

The website lists five prerequisites:

• proficiency in Python, plus software engineering skills;
• experience with deep learning and systems optimization, familiarity with PyTorch and an understanding of basic systems concepts such as the memory hierarchy;
• college calculus and linear algebra, including the ability to read matrix and vector notation;
• basic probability and statistics, such as Gaussian distributions, means and standard deviations;
• machine learning fundamentals equivalent to one of Stanford's CS221, CS229, CS230, CS124 or CS224N.

The website specifically warns that the assignments provide almost no scaffolding and require at least an order of magnitude more code than other courses. CS336 is a 5-unit course, and Stanford Online lists the expected commitment as 10 weeks at 20–25 hours per week.

You have to arrange your own compute. The website lists public prices for a single B200 GPU as of March 28, 2026:

• Modal, the course sponsor: $6.25 per hour, plus $30 of free credit each month;
• Lambda: $6.69;
• RunPod: $4.99;
• Nebius: $5.50 ($3.05 for preemptible instances);
• Together: $7.49, with a minimum rental of 8 GPUs.

The website recommends getting your code correct on CPU first. After that, use GPUs for the training in Assignments 1, 4 and 5 and for the GPU benchmarking in Assignment 2, with the number of GPUs recommended for each assignment. Assignment 4 needs the most compute. Its README states that final training uses 8 B200s, which comes to about $50 per hour at Modal's list price. The actual training time depends on the assignment handout (to be verified). Assignment 3's hosted training API requires a Stanford student ID, so learners who are not enrolled must run the training scripts themselves by following the README and pay for that compute too. The course does not estimate the total cost of the whole course, so the total is still to be verified. These prices may also change as the platforms update them.

A compute cluster equipped with NVIDIA Tesla GPUs, shown only to illustrate GPU compute. These are not the B200s used for the CS336 assignments. Photo: ChrisDag. Source: Wikimedia Commons. License: CC BY 2.0.

The free public website, recordings and assignments do not come with a certificate. The website says enrolled students submit through Gradescope, and it does not mention any grading option for self-learners. If you need a formal credential, Stanford Online offers CS336 as a graduate course that you can apply to. Tuition is $7,875, although the page notes that it may change. The course is worth 5 units, and the credential is a Stanford University transcript. The course can count toward the Artificial Intelligence Graduate Certificate program, and the class dates for the 2026 offering are March 30 to June 10. As of October 22, 2026, the page shows that enrollment is currently closed and lists the spring quarter as "tentative". Because of compute and workload limits, enrollment uses a waitlist. Applicants must submit a non-degree application and a course-specific application. They must also have a bachelor's degree, an undergraduate GPA of at least 3.0 and credit for CS224N or a similar natural language processing course. The page does not state the application deadline or eligible regions for the next offering, so these details are still to be verified.

Suggested Self-Study Order

This site compiled the order below from the course schedule and each assignment's README. It is a suggestion, not an official course requirement.

1. Cover the prerequisites first: If you are not yet comfortable with PyTorch, practice writing tensor operations, using autograd and building a complete training loop before you start CS336. Karpathy's eight-lecture neural network series, which this site has covered before, can serve as a bridge.
2. Lectures 1–4 + Assignment 1: Clone the repository, manage the environment with uv and run uv run pytest. At first, every test fails with NotImplementedError. Connect your own implementations in tests/adapters.py and get the tests passing one at a time. Once the code is correct on CPU, train a small model on TinyStories.
3. Lectures 5–8 + Assignment 2: From this point on, you need a GPU. The Assignment 2 repository includes the official reference implementation of Assignment 1, so you can continue even if you have not fully finished Assignment 1. You can also watch Lecture 10 on inference during this stage.
4. Lectures 9 and 11 + Assignment 3: Focus on understanding how scaling laws are fitted and extrapolated. Learners who are not enrolled cannot use the hosted training API, so estimate the compute cost of running the training yourself before you start.
5. Lectures 12–14 + Assignment 4: Start by downloading a small number of files in offline mode and building the filtering and deduplication pipeline. The README notes that you can reduce the number of WET files downloaded (about 2,500 by default). On a limited budget, you can work at a smaller scale instead of reproducing the 8-GPU final training run.
6. Lectures 15–17 + Assignment 5: Complete the supervised fine-tuning and reinforcement learning parts. If you still have time and energy, do the optional safety alignment part (DPO and similar methods). You can leave the two guest lectures until the end.

The official pace covers all the material in 10 weeks. If you are studying while working, you can stretch the timeline to fit the time you actually have each week. Work through one assignment at a time, and move on to the next group of lectures after you finish each assignment.

References

1. Stanford CS336 | Language Modeling from Scratch, course website, accessed October 22, 2026

2. Stanford CS336: Language Modeling from Scratch | Spring 2026, YouTube lecture recording playlist

3. stanford-cs336, GitHub organization page

4. stanford-cs336/lectures, lecture materials repository

5. stanford-cs336/assignment1-basics

6. stanford-cs336/assignment2-systems

7. stanford-cs336/assignment3-scaling

8. stanford-cs336/assignment4-data

9. stanford-cs336/assignment5-alignment

10. stanford-cs336.github.io commit history, course website repository

11. Language Modeling from Scratch | Stanford Online, accessed October 22, 2026

12. Smart Academy · Your own academy for the intelligent era

13. Compiled and published by Scholay, the one-stop academic platform. If this content infringes your rights, please let us know and it will be taken down.

14. Also published on EnterScholar, a community for scholars in the intelligent era