Spaces:
Sleeping
Sleeping
Report for cardiffnlp/twitter-roberta-base-offensive
#192
by
giskard-bot
- opened
Hi Team,
This is a report from Giskard Bot Scan 🐢.
We have identified 6 potential vulnerabilities in your model based on an automated scan.
This automated analysis evaluated the model on the dataset tweet_eval (subset offensive, split test).
You can find a full version of scan report here.
👉Robustness issues (3)
When feature “text” is perturbed with the transformation “Transform to uppercase”, the model changes its prediction in 7.86% of the cases. We expected the predictions not to be affected by this transformation.
| Level | Metric | Transformation | Deviation |
|---|---|---|---|
| medium 🟡 | Fail rate = 0.079 | Transform to uppercase | 67/852 tested samples (7.86%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201🔍✨Examples
| text | Transform to uppercase(text) | Original prediction | Prediction after perturbation | |
|---|---|---|---|---|
| 10 | #Kavanaugh is a disciple of Anthony Kennedy, who 🎁ed liberals w/3 unconstitutional opinions! He's Trump's con to con conservatives to 🤔ing #SCOTUS is moving right! Were @user smart, they'd do everything to get BK the 🏁. But they won't! Idiots! #TuesdayThoughts | #KAVANAUGH IS A DISCIPLE OF ANTHONY KENNEDY, WHO 🎁ED LIBERALS W/3 UNCONSTITUTIONAL OPINIONS! HE'S TRUMP'S CON TO CON CONSERVATIVES TO 🤔ING #SCOTUS IS MOVING RIGHT! WERE @USER SMART, THEY'D DO EVERYTHING TO GET BK THE 🏁. BUT THEY WON'T! IDIOTS! #TUESDAYTHOUGHTS | offensive (p = 0.74) | non-offensive (p = 0.54) |
| 32 | #BeckyLynch is beautiful one of the few women in wrestling that absolutely need no work done at all. She's absolutely beautiful just the way she is. Doesn't need giant boobs or a fake booty. @user is just simply amazing #HIAC | #BECKYLYNCH IS BEAUTIFUL ONE OF THE FEW WOMEN IN WRESTLING THAT ABSOLUTELY NEED NO WORK DONE AT ALL. SHE'S ABSOLUTELY BEAUTIFUL JUST THE WAY SHE IS. DOESN'T NEED GIANT BOOBS OR A FAKE BOOTY. @USER IS JUST SIMPLY AMAZING #HIAC | offensive (p = 0.61) | non-offensive (p = 0.52) |
| 63 | @user @user Awwww she is so stinking cute! How old is she now? | @USER @USER AWWWW SHE IS SO STINKING CUTE! HOW OLD IS SHE NOW? | offensive (p = 0.73) | non-offensive (p = 0.50) |
When feature “text” is perturbed with the transformation “Transform to title case”, the model changes its prediction in 5.36% of the cases. We expected the predictions not to be affected by this transformation.
| Level | Metric | Transformation | Deviation |
|---|---|---|---|
| medium 🟡 | Fail rate = 0.054 | Transform to title case | 46/858 tested samples (5.36%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201🔍✨Examples
| text | Transform to title case(text) | Original prediction | Prediction after perturbation | |
|---|---|---|---|---|
| 2 | ...if you want more shootings and more death, then listen to the ACLU, Black Lives Matter, or Antifa. If you want public safety, then listen to the police professionals who have been studying this for 35 years."" -AG Jeff Sessions | ...If You Want More Shootings And More Death, Then Listen To The Aclu, Black Lives Matter, Or Antifa. If You Want Public Safety, Then Listen To The Police Professionals Who Have Been Studying This For 35 Years."" -Ag Jeff Sessions | non-offensive (p = 0.60) | offensive (p = 0.54) |
| 9 | #RAP is a form of ART! Used to express yourself freely. It does not gv the green light or excuse the behavior of acting like an animal! She is not in the streets of the BX where violence is a way of living. Elevate yourself boo and get on @user level for longevity! #QUEEN👑 | #Rap Is A Form Of Art! Used To Express Yourself Freely. It Does Not Gv The Green Light Or Excuse The Behavior Of Acting Like An Animal! She Is Not In The Streets Of The Bx Where Violence Is A Way Of Living. Elevate Yourself Boo And Get On @User Level For Longevity! #Queen👑 | non-offensive (p = 0.51) | offensive (p = 0.56) |
| 25 | #Jenelle wants the world to know she is in a bikini. Oh, and to pray for NC. 😒 | #Jenelle Wants The World To Know She Is In A Bikini. Oh, And To Pray For Nc. 😒 | non-offensive (p = 0.80) | offensive (p = 0.51) |
When feature “text” is perturbed with the transformation “Add typos”, the model changes its prediction in 5.09% of the cases. We expected the predictions not to be affected by this transformation.
| Level | Metric | Transformation | Deviation |
|---|---|---|---|
| medium 🟡 | Fail rate = 0.051 | Add typos | 41/805 tested samples (5.09%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201🔍✨Examples
| text | Add typos(text) | Original prediction | Prediction after perturbation | |
|---|---|---|---|---|
| 10 | #Kavanaugh is a disciple of Anthony Kennedy, who 🎁ed liberals w/3 unconstitutional opinions! He's Trump's con to con conservatives to 🤔ing #SCOTUS is moving right! Were @user smart, they'd do everything to get BK the 🏁. But they won't! Idiots! #TuesdayThoughts | #Kavanaug us a disciple of Antbony Kennecy, woh 🎁ed liberals w/3 unconstitutional opinions! He's Trump's con ti con conservtaives to 🤔ing #SVOTUS s moving right! Were @user smart, thry'd do everything to get BK th 🏁. But they won't! diots! #TuexdayThouthts | offensive (p = 0.74) | non-offensive (p = 0.59) |
| 15 | And apparently I'm committed to going to a new level since I used the key. Well FUCK. Curiosity killed the Cat(hy) | And apparently I'm committed to going to a new level since I ssd the key. Well HFCK. Curiosity killd the Cat(hy) | offensive (p = 0.85) | non-offensive (p = 0.75) |
| 69 | @user Gun control? I'm wondering who was in control of all those bodies that got suicided because of you. You need to be controlled. 24/7/365 #GITMO THE PAIN IS COMING. #MAGA #QAnon #Benghazi Hope they televise your demise. | @user Gun conyrol? I'm wondwering who was in control of all those bodis sthat got suicided because of you. You need to be controlled. 24/7/365 #GITMO THE PAIB IS COMING. #MAGA #QAnon #Benguazi Uope they televise your emise. | offensive (p = 0.61) | non-offensive (p = 0.58) |
👉Performance issues (3)
For records in the dataset where text contains "antifa", the Recall is 50.0% lower than the global Recall.
| Level | Data slice | Metric | Deviation |
|---|---|---|---|
| major 🔴 | text contains "antifa" |
Recall = 0.333 | -50.00% than global |
Taxonomy
avid-effect:performance:P0204🔍✨Examples
| text | label | Predicted label |
|
|---|---|---|---|
| 1 | @user @user @user I got in a pretty deep debate with my friend and she told me that latinos for Trump and blacks for Trump were paid supporters 😂 then I said you mean antifa are paid domestic terrorist and she said No they are anti-fascist then I said they are the fascist are you kidding me?! | non-offensive | offensive (p = 0.65) |
| 148 | #Antifa protest in #Athens for the 5 years after the murder of Pavlos Fyssas by Golden Dawn. Their voters are complicit to their crimes. #KillahP | offensive | non-offensive (p = 0.55) |
| 189 | @user @user @user Fascism was also against liberals as is Antifa according to your meme. | offensive | non-offensive (p = 0.80) |
For records in the dataset where text contains "conservatives", the Recall is 31.82% lower than the global Recall.
| Level | Data slice | Metric | Deviation |
|---|---|---|---|
| major 🔴 | text contains "conservatives" |
Recall = 0.455 | -31.82% than global |
Taxonomy
avid-effect:performance:P0204🔍✨Examples
| text | label | Predicted label |
|
|---|---|---|---|
| 35 | #ConstitutionDay It's very odd for the alt right conservatives to say that we are ruining the constitution just because we want #GunControlNow but they are the ones ruining the constitution getting upset because foreigners are coming to this land who are not White wanting to live | offensive | non-offensive (p = 0.58) |
| 62 | #Conservatives We mus pray for liberals. They trooly kno knot watt they do. I was brot up socialist democrat, by my wonderful, but, ignorant parents. I served inda military. I learnd I can serve Christ, too. Eye reel -eyesed CONSERVATIVISM was more closely aligned w God. | offensive | non-offensive (p = 0.68) |
| 80 | #BlueWave #DumpTrump …give my love 2, or cooperate/compromise w/, conservatives/Trump supporters who constantly vilify marginalized groups, who constantly attack women’s rights 2 make choices about their own bodies,… | offensive | non-offensive (p = 0.62) |
For records in the dataset where text contains "trump", the Balanced Accuracy is 10.74% lower than the global Balanced Accuracy.
| Level | Data slice | Metric | Deviation |
|---|---|---|---|
| major 🔴 | text contains "trump" |
Balanced Accuracy = 0.714 | -10.74% than global |
Taxonomy
avid-effect:performance:P0204🔍✨Examples
| text | label | Predicted label |
|
|---|---|---|---|
| 1 | @user @user @user I got in a pretty deep debate with my friend and she told me that latinos for Trump and blacks for Trump were paid supporters 😂 then I said you mean antifa are paid domestic terrorist and she said No they are anti-fascist then I said they are the fascist are you kidding me?! | non-offensive | offensive (p = 0.65) |
| 80 | #BlueWave #DumpTrump …give my love 2, or cooperate/compromise w/, conservatives/Trump supporters who constantly vilify marginalized groups, who constantly attack women’s rights 2 make choices about their own bodies,… | offensive | non-offensive (p = 0.62) |
| 122 | #America ... tear down that #Wall! #tcot #partisanship #Trump #thewall #Borderwall #liberty #civilsociety #think #Conservatives #Democrats #Progressives #liberals #Independent #libertarians #GOP #DNC #CriticalThinking | offensive | non-offensive (p = 0.90) |
Checkout out the Giskard Space and Giskard Documentation to learn more about how to test your model.
Disclaimer: it's important to note that automated scans may produce false positives or miss certain vulnerabilities. We encourage you to review the findings and assess the impact accordingly.