The evaluation of the quality model was replicated and extended in T360576. This task focuses on the replication of the training of the model (so taking a step backward): https://public-paws.wmcloud.org/User:Isaac%20(WMF)/Quality/Quality_Model_Training.ipynb
== 1. Getting started
[x] Fork [[https://public-paws.wmcloud.org/User:Isaac%20(WMF)/Quality/Quality_Model_Training.ipynb|the notebook]] ([[https://wikitech.wikimedia.org/wiki/PAWS/Getting_started_with_PAWS#Fork|directions]]) and add to your PAWS directory
[x] Try to understand the existing code -- make sure you know what each function is accomplishing and ask questions where it's less clear. Some of the code will be familiar from the model evaluation but some will be new.
== 2. Add in Chinese support
Incorporate the function for including Chinese Wikipedia data that was created in T360576:
[x] The translation map code can be added with the other language functions in the [[https://public-paws.wmcloud.org/User:Isaac%20(WMF)/Quality/Quality_Model_Training.ipynb#Fetching-data|fetching data section]]
[] zhwiki should be added to the [[https://public-paws.wmcloud.org/User:Isaac%20(WMF)/Quality/Quality_Model_Training.ipynb#Load-in-data|data-loading configuration]] so it's included in the model training
[] You'll also have to add in the category/media prefixes used in Chinese Wikipedia which can be found [[https://github.com/wikimedia/research-api-endpoint-template/blob/quality-article/model/wsgi.py#L299|here]] and [[https://github.com/wikimedia/research-api-endpoint-template/blob/quality-article/model/wsgi.py#L554|here]] and are used in the `Feature extraction` section of the notebook
== 3. Replication
[] Run the code from start-to-finish!
[] In markdown at the top of the notebook, write a summary that includes your new model coefficients and how similar they are to the [[https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/Language-agnostic_Wikipedia_article_quality#Implementation|original feature weights]] as well as what I got when I created the notebook (I never made a summary myself but you can see them in the cell outputs towards the end of the notebook). NOTE: the negative coefficient that I got for categories is non-desirable but not super surprising. We can discuss when I get back.
== 4. Optional explorations
[] I made basic scatter plots of predicted scores and true scores (similar to the evaluation). The color-coding (blue - yellow) leaves it highly unclear as to what color maps to which Wikipedia language edition. If you have time, you could see how to generate a clearer legend that gives each language edition a clear color category that is labeled in a legend. Probably something like [[https://seaborn.pydata.org/generated/seaborn.scatterplot.html|this function]] -- seaborn is a non-standard library but you can install it like I did for `wmpaws` at the top of the notebook.
[] We identified that Nepalese Wikipedia (newiki) [[https://extloc.toolforge.org/extensions/PageAssessments|also uses PageAssessments]]. We could also incorporate that into training/evaluation by following the steps for Chinese Wikipedia (T360576). One thing to be aware of: Nepalese Wikipedia does not seem to have a content assessment rubric however, so this might be a bit trickier. Hopefully Google Translate + the labels in the database might be sufficient to get a reasonable dataset. It's also a smaller language so I'd consider extending the sample dates to be more than 28 days (possibly do a full year).