giskardai/giskard-evaluator · Report for cardiffnlp/twitter-roberta-base-offensive

Hi Team,

This is a report from Giskard Bot Scan 🐢.

We have identified 16 potential vulnerabilities in your model based on an automated scan.

This automated analysis evaluated the model on the dataset tweet_eval (subset offensive, split train).

👉Robustness issues (3)

When feature “text” is perturbed with the transformation “Transform to uppercase”, the model changes its prediction in 12.2% of the cases. We expected the predictions not to be affected by this transformation.

Level	Data slice	Metric	Deviation
major 🔴	—	Fail rate = 0.122	122/1000 tested samples (12.2%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201

🔍✨Examples

	text	Transform to uppercase(text)	Original prediction	Prediction after perturbation
3031	#Liberals had some sort of mutual masturbation awards show. This is me still not caring. @user @user @user @user	#LIBERALS HAD SOME SORT OF MUTUAL MASTURBATION AWARDS SHOW. THIS IS ME STILL NOT CARING. @USER @USER @USER @USER	offensive (p = 0.64)	non-offensive (p = 0.84)
5121	@user His name is Coco! He is a Pomeranian. He's actually old. He's is 12. The most loyal dog I have even known!	@USER HIS NAME IS COCO! HE IS A POMERANIAN. HE'S ACTUALLY OLD. HE'S IS 12. THE MOST LOYAL DOG I HAVE EVEN KNOWN!	non-offensive (p = 0.80)	offensive (p = 0.65)
10544	@user I`m with you JR.She is nothing.Don`t waste your time on human waste!	@USER I`M WITH YOU JR.SHE IS NOTHING.DON`T WASTE YOUR TIME ON HUMAN WASTE!	offensive (p = 0.55)	non-offensive (p = 0.69)

When feature “text” is perturbed with the transformation “Add typos”, the model changes its prediction in 8.2% of the cases. We expected the predictions not to be affected by this transformation.

Level	Data slice	Metric	Deviation
medium 🟡	—	Fail rate = 0.082	82/1000 tested samples (8.2%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201

🔍✨Examples

	text	Add typos(text)	Original prediction	Prediction after perturbation
8437	#guncontrol itself specifically targets minorities, @user The entire reason gun control"" exists is pure, unadulterated racism - bigots like you wanted to keep the recently-freed slaves disarmed and subordinate. Thankfully we've grown from there.	#guncontril itself specifically targtes minoriues, @user The entire reason gun control"" existw is pure, unadulterated racism - bigots like you wanted to mkeep tye recently-free dslafves disarmed an dsubordintae.. Thankfully we'be grpwn fro mthere.	offensive (p = 0.58)	non-offensive (p = 0.52)
11540	@user When do you post on conservatives declaring Ford a liar without evidence?	@user When do you post on vconservatives declaring Ford a kiar without evidence?	offensive (p = 0.78)	non-offensive (p = 0.78)
147	@user @user He is such a showman! And those pipes😍 the man can SING	@user @user He is such a showmsan! Xnd those piped😍 t ema ncan ZING	offensive (p = 0.51)	non-offensive (p = 0.86)

When feature “text” is perturbed with the transformation “Transform to title case”, the model changes its prediction in 7.8% of the cases. We expected the predictions not to be affected by this transformation.

Level	Data slice	Metric	Deviation
medium 🟡	—	Fail rate = 0.078	78/1000 tested samples (7.8%) changed prediction after perturbation

Taxonomy

avid-effect:performance:P0201

🔍✨Examples

	text	Transform to title case(text)	Original prediction	Prediction after perturbation
3031	#Liberals had some sort of mutual masturbation awards show. This is me still not caring. @user @user @user @user	#Liberals Had Some Sort Of Mutual Masturbation Awards Show. This Is Me Still Not Caring. @User @User @User @User	offensive (p = 0.64)	non-offensive (p = 0.79)
10185	Baaaaa baaaaa...that's the American sheeple who prove Roger Goodell right: The #NFL CAN do whatever it wants and you'll take it because you'll NEVER give up the NFL. Baaaaa baaaaa. @user #QAnon #MAGA #WWG1WGA #FreeAlexJones	Baaaaa Baaaaa...That'S The American Sheeple Who Prove Roger Goodell Right: The #Nfl Can Do Whatever It Wants And You'Ll Take It Because You'Ll Never Give Up The Nfl. Baaaaa Baaaaa. @User #Qanon #Maga #Wwg1Wga #Freealexjones	offensive (p = 0.62)	non-offensive (p = 0.59)
2192	@user @user @user @user @user Only liberals can take an athlete berating an official and turn it into sexism and a sjw issue. Liberals ruin everything by putting politics into everything.	@User @User @User @User @User Only Liberals Can Take An Athlete Berating An Official And Turn It Into Sexism And A Sjw Issue. Liberals Ruin Everything By Putting Politics Into Everything.	non-offensive (p = 0.52)	offensive (p = 0.58)

👉Performance issues (6)

For records in the dataset where text contains "maga", the Recall is 20.67% lower than the global Recall.

Level	Data slice	Metric	Deviation
major 🔴	`text` contains "maga"	Recall = 0.633	-20.67% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
115	@user @user @user How about your outrage over this? Do something so you can be credible instead of being perceived as demagogues partisans #MAGA	offensive	non-offensive (p = 0.58)
263	Sean Penn: #MeToo Movement Seeks to ‘Divide Men and Women’ <— Gotta say… I stand w/ Jeff Spicoli on this one. #MAGA #WWG1WGA #ConfirmKavanaugh	offensive	non-offensive (p = 0.79)
284	@user @user Silly Joe! I’m a MAGA Of Society 💯🇺🇸❤️🇺🇸❤️🇺🇸... Feel That👊👊	non-offensive	offensive (p = 0.64)

For records in the dataset where text contains "conservatives", the Recall is 19.71% lower than the global Recall.

Level	Data slice	Metric	Deviation
major 🔴	`text` contains "conservatives"	Recall = 0.641	-19.71% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
10	@user Antifa would burn a Conservatives house down and CNN would be there lighting the torches & throwing gas on the flames.	offensive	non-offensive (p = 0.58)
98	@user #Conservatives #Republicans are #TRAITORS to #America	offensive	non-offensive (p = 0.82)
493	@user Conservatives who quit rather than fight betray every claim that cons are patriots"" & ""will fight for our Constitution."" It reveals extremism & defeatism at once. Dems KNOW they will win if they stick together. History PROVES it.""	offensive	non-offensive (p = 0.78)

For records in the dataset where text contains "antifa", the Precision is 9.72% lower than the global Precision.

Level	Data slice	Metric	Deviation
medium 🟡	`text` contains "antifa"	Precision = 0.716	-9.72% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
10	@user Antifa would burn a Conservatives house down and CNN would be there lighting the torches & throwing gas on the flames.	offensive	non-offensive (p = 0.58)
156	@user A liberal anarchist coup d'etat gang runs rampant & a GOP AG does nothing. Liberals prove to be the biggest hate group in USA. The leftstream media cheers them on while excoriating duly elected President of the United States 24/7. ANTIFA openly advocates killing cops. Silence	non-offensive	offensive (p = 0.65)
259	AntiFa fascists and Socialists are one and the same......time to respond !	offensive	non-offensive (p = 0.57)

For records in the dataset where text contains "liberals", the Precision is 8.59% lower than the global Precision.

Level	Data slice	Metric	Deviation
medium 🟡	`text` contains "liberals"	Precision = 0.725	-8.59% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
156	@user A liberal anarchist coup d'etat gang runs rampant & a GOP AG does nothing. Liberals prove to be the biggest hate group in USA. The leftstream media cheers them on while excoriating duly elected President of the United States 24/7. ANTIFA openly advocates killing cops. Silence	non-offensive	offensive (p = 0.65)
159	@user @user Why don't liberals understand that an allegation of drunk teenagers who don't have sex is not a federal crime? Golly Moses- What is wrong with those people?	non-offensive	offensive (p = 0.66)
338	@user @user Of course they are. She either agrees to testify on Mon or they vote on Friday and life moves on until the next obstruction from the lunatic liberals	non-offensive	offensive (p = 0.56)

For records in the dataset where text_length(text) >= 117.500, the Recall is 6.88% lower than the global Recall.

Level	Data slice	Metric	Deviation
medium 🟡	`text_length(text)` >= 117.500	Recall = 0.743	-6.88% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	text_length(text)	label	Predicted `label`
0	@user Bono... who cares. Soon people will understand that they gain nothing from following a phony celebrity. Become a Leader of your people instead or help and support your fellow countrymen.	192	non-offensive	offensive (p = 0.55)
10	@user Antifa would burn a Conservatives house down and CNN would be there lighting the torches & throwing gas on the flames.	128	offensive	non-offensive (p = 0.58)
23	@user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user @user He also clarified that there is the water birth when we are physically born. Then there is the spiritual birth which is obviously when we spiritually die to our sinful flesh and be reborn sinless after being cleansed by His blood at salvation when our sin debt is paid! Praise God	556	offensive	non-offensive (p = 0.69)

For records in the dataset where text contains "don", the Recall is 6.6% lower than the global Recall.

Level	Data slice	Metric	Deviation
medium 🟡	`text` contains "don"	Recall = 0.745	-6.60% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
112	@user @user @user Do it with quran and then we will see how much atheist and brave she is... hindus figured it out that this type of attention grabbing big boss wanna be people are not worth it... Go live in your 2 minutes Prime Time covered life ... Lord krishna and his followers don't care	offensive	non-offensive (p = 0.54)
159	@user @user Why don't liberals understand that an allegation of drunk teenagers who don't have sex is not a federal crime? Golly Moses- What is wrong with those people?	non-offensive	offensive (p = 0.66)
375	@user Bye bye Flake you Flake! Don’t let the Senate door hit you in the behind on your way out. @user @user @user @user	offensive	non-offensive (p = 0.52)

👉Underconfidence issues (7)

For records in your dataset where text contains "don", we found a significantly higher number of underconfident predictions (23 samples, corresponding to 3.20% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
major 🔴	`text` contains "don"	Underconfidence rate = 0.032	+89.91% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
520	@user The Dems playing a con game. It must be Antifa don’t you think.	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)
3441	@user @user @user @user @user BS. No we don't support gun control. We support #2A that says SHALL NOT BE INFRINGED. Go read some history and find out why. You are part of the reason.	offensive	non-offensive (p = 0.50)
			offensive (p = 0.50)
2484	@user @user @user @user @user @user You are a child. Don’t even talk to me. Lmao.	non-offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)

For records in your dataset where text contains "liberals", we found a significantly higher number of underconfident predictions (35 samples, corresponding to 2.82% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
major 🔴	`text` contains "liberals"	Underconfidence rate = 0.028	+67.06% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
10284	20 Crazy Things Liberals Blame Trump For via @user Unhinged blame everything on our President.	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)
8538	@user @user Tommy Robinson speaks the truth and finds himself in jail? Le Penspeaks the truth and now she needs a psychiatric evaluation? America wake up! Vote Democrat and this is your future and your kids future! Democrats/liberals support muslim extremists !	non-offensive	non-offensive (p = 0.50)
			offensive (p = 0.50)
8551	@user Maybe we need more gun control after all - liberals should not be allowed to possess them!	non-offensive	non-offensive (p = 0.50)
			offensive (p = 0.50)

For records in your dataset where text contains "antifa", we found a significantly higher number of underconfident predictions (28 samples, corresponding to 2.68% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
major 🔴	`text` contains "antifa"	Underconfidence rate = 0.027	+59.15% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
520	@user The Dems playing a con game. It must be Antifa don’t you think.	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)
10981	@user Why is Antifa upset w/ “prof.” Blasé-Ford? Antifa is the group that harrasses & threatens people at their homes.	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)
5348	@user Please show me your post where you said the same thing about conservatives targeted by Antifa and other leftist friends of yours 🙄 #Hypocrite	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)

For records in your dataset where text contains "conservatives", we found a significantly higher number of underconfident predictions (21 samples, corresponding to 2.53% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
major 🔴	`text` contains "conservatives"	Underconfidence rate = 0.025	+49.99% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	label	Predicted `label`
5348	@user Please show me your post where you said the same thing about conservatives targeted by Antifa and other leftist friends of yours 🙄 #Hypocrite	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)
6667	@user @user @user Conservatives are whining"". There, I fixed the typo for you.""	offensive	offensive (p = 0.50)
			non-offensive (p = 0.50)
11156	@user The reason people are tired of this is because you only focus on the conservatives. Ford has treated people poorly and the constitutional crisis is abominable. His treatment of the women in the legislature is despicable.	non-offensive	non-offensive (p = 0.50)
			offensive (p = 0.50)

For records in your dataset where text_length(text) >= 79.500, we found a significantly higher number of underconfident predictions (157 samples, corresponding to 2.17% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
major 🔴	`text_length(text)` >= 79.500	Underconfidence rate = 0.022	+28.88% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	text_length(text)	label	Predicted `label`
2757	. @user tells us she is an unserious attention seeker #iphone @user @user @user	81	offensive	offensive (p = 0.50)
				non-offensive (p = 0.50)
11638	@user @user @user @user @user @user @user @user @user @user @user Yea cuz mushroom penis syndrome is definitely an impeachable offense 😂 #MAGA	142	offensive	non-offensive (p = 0.50)
				offensive (p = 0.50)
10284	20 Crazy Things Liberals Blame Trump For via @user Unhinged blame everything on our President.	95	offensive	offensive (p = 0.50)
				non-offensive (p = 0.50)

For records in your dataset where avg_word_length(text) >= 4.009, we found a significantly higher number of underconfident predictions (188 samples, corresponding to 1.94% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
medium 🟡	`avg_word_length(text)` >= 4.009	Underconfidence rate = 0.019	+15.01% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	avg_word_length(text)	label	Predicted `label`
2757	. @user tells us she is an unserious attention seeker #iphone @user @user @user	4.71429	offensive	offensive (p = 0.50)
				non-offensive (p = 0.50)
436	@user He is reprehensible. Him and Durbin.	5.14286	non-offensive	non-offensive (p = 0.50)
				offensive (p = 0.50)
11638	@user @user @user @user @user @user @user @user @user @user @user Yea cuz mushroom penis syndrome is definitely an impeachable offense 😂 #MAGA	5.21739	offensive	non-offensive (p = 0.50)
				offensive (p = 0.50)

For records in your dataset where avg_whitespace(text) < 0.189, we found a significantly higher number of underconfident predictions (182 samples, corresponding to 1.91% of the predictions in the data slice).

Level	Data slice	Metric	Deviation
medium 🟡	`avg_whitespace(text)` < 0.189	Underconfidence rate = 0.019	+13.00% than global

Taxonomy

avid-effect:performance:P0204

🔍✨Examples

	text	avg_whitespace(text)	label	Predicted `label`
11293	@user Oh my god🤮🤮🤮🤮	0.157895	non-offensive	offensive (p = 0.50)
				non-offensive (p = 0.50)
2757	. @user tells us she is an unserious attention seeker #iphone @user @user @user	0.185185	offensive	offensive (p = 0.50)
				non-offensive (p = 0.50)
436	@user He is reprehensible. Him and Durbin.	0.142857	non-offensive	non-offensive (p = 0.50)
				offensive (p = 0.50)

Checkout out the Giskard Space and 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.