Hi Team,
This is a report from Giskard Bot Scan 🐢.
We have identified 8 potential vulnerabilities in your model based on an automated scan.
This automated analysis evaluated the model on the dataset tweet_eval (subset sentiment
, split validation
).
You can find a full version of scan report here.
👉Robustness issues (5)
When feature “text” is perturbed with the transformation “Add typos”, the model changes its prediction in 15.1% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
major 🔴 |
Fail rate = 0.151 |
Add typos |
151/1000 tested samples (15.1%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Add typos(text) |
Original prediction |
Prediction after perturbation |
1635 |
"on Black Friday i always thought Kendrick said ""Coney Island!!"" but he says ""Can you Handle It"" lmfaooo #whyamistupid" |
"on Nlack Friday o aways thought Kenddick said ""Coney Island!!"" bjut he says ""Can you Handle It"" lmfaooo #whyamistupid" |
neutral (p = 0.46) |
negative (p = 0.54) |
1254 |
Hillary's campaign now reset for the 4th time. Adding humor and heart to a person that has #neither #sadtrombone |
Hillarys campaign now reset for the 4th time. Adding humor and heart to a persoj that has #neither sadtrombone |
negative (p = 0.62) |
neutral (p = 0.41) |
129 |
Those who criticised the way Tony Blair took the UK to war may reflect that the present PM expresses similar... |
Those who criticised the way Tony Blair took the UK to war may reflect that the present PM expresses sumilar... |
neutral (p = 0.51) |
negative (p = 0.53) |
When feature “text” is perturbed with the transformation “Transform to uppercase”, the model changes its prediction in 14.7% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
major 🔴 |
Fail rate = 0.147 |
Transform to uppercase |
147/1000 tested samples (14.7%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Transform to uppercase(text) |
Original prediction |
Prediction after perturbation |
1666 |
"If it ain't broke don't fix it, why move kris Bryant up to 3rd when he's hitting as good as he has all season at 5" |
"IF IT AIN'T BROKE DON'T FIX IT, WHY MOVE KRIS BRYANT UP TO 3RD WHEN HE'S HITTING AS GOOD AS HE HAS ALL SEASON AT 5" |
neutral (p = 0.65) |
negative (p = 0.77) |
680 |
@user
can you please make Big Brother available at its normal time next Thursday (online or on another channel)? Thank you. |
@USER
CAN YOU PLEASE MAKE BIG BROTHER AVAILABLE AT ITS NORMAL TIME NEXT THURSDAY (ONLINE OR ON ANOTHER CHANNEL)? THANK YOU. |
neutral (p = 0.55) |
positive (p = 0.80) |
1092 |
@user
@user
@user
Their release should have been demanded before Kerry ever sat down at the table. |
@USER
@USER
@USER
THEIR RELEASE SHOULD HAVE BEEN DEMANDED BEFORE KERRY EVER SAT DOWN AT THE TABLE. |
negative (p = 0.61) |
neutral (p = 0.56) |
When feature “text” is perturbed with the transformation “Transform to title case”, the model changes its prediction in 9.2% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
medium 🟡 |
Fail rate = 0.092 |
Transform to title case |
92/1000 tested samples (9.2%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Transform to title case(text) |
Original prediction |
Prediction after perturbation |
1242 |
the most important thing madonna has ever said is " don't go for 2nd best " |
The Most Important Thing Madonna Has Ever Said Is " Don'T Go For 2Nd Best " |
neutral (p = 0.49) |
positive (p = 0.53) |
1636 |
@user
They're actually going venue shopping tomorrow! They're checking out Grand Bend and surrounding areas (ie. St. Mary's)! |
@User
They'Re Actually Going Venue Shopping Tomorrow! They'Re Checking Out Grand Bend And Surrounding Areas (Ie. St. Mary'S)! |
positive (p = 0.63) |
neutral (p = 0.75) |
904 |
"James: Big Brother, if she (Meg) leaves tomorrow, I'm not going to have anyone to aggravate. #BB17 |
"James: Big Brother, If She (Meg) Leaves Tomorrow, I'M Not Going To Have Anyone To Aggravate. #Bb17 |
negative (p = 0.51) |
neutral (p = 0.56) |
When feature “text” is perturbed with the transformation “Punctuation Removal”, the model changes its prediction in 8.2% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
medium 🟡 |
Fail rate = 0.082 |
Punctuation Removal |
82/1000 tested samples (8.2%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Punctuation Removal(text) |
Original prediction |
Prediction after perturbation |
1489 |
Curtis Painter...we have a chance again! Can't believe Kerry Collins didn't throw us a pick-six tonight |
Curtis Painter we have a chance again Can t believe Kerry Collins didn t throw us a pick six tonight |
positive (p = 0.69) |
neutral (p = 0.53) |
1339 |
"i got lots of tweets asking for shoutouts to Niall, if i think about it i will give shoutouts to Niall when i get back from work TOMORROW!!" |
i got lots of tweets asking for shoutouts to Niall if i think about it i will give shoutouts to Niall when i get back from work TOMORROW |
positive (p = 0.69) |
neutral (p = 0.54) |
1952 |
@user
@user
Yellow journalism. But you know? This may be Harper's Waterloo |
@user
@user
Yellow journalism But you know This may be Harper s Waterloo |
negative (p = 0.56) |
neutral (p = 0.67) |
When feature “text” is perturbed with the transformation “Transform to lowercase”, the model changes its prediction in 5.2% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
medium 🟡 |
Fail rate = 0.052 |
Transform to lowercase |
52/1000 tested samples (5.2%) changed prediction after perturbation |
Taxonomy
avid-effect:performance:P0201
🔍✨Examples
|
text |
Transform to lowercase(text) |
Original prediction |
Prediction after perturbation |
77 |
@user
seriously! itunes puts like an entire minute as a preview so 20 seconds is nothing. AND I KNOW! it needs to be monday ASAP! |
@user
seriously! itunes puts like an entire minute as a preview so 20 seconds is nothing. and i know! it needs to be monday asap! |
negative (p = 0.46) |
neutral (p = 0.48) |
756 |
NIKE EMPLOYEE'S: If anyone want to work tomorrow at 5am call!!!!!!!!!!!!!!!!!! |
nike employee's: if anyone want to work tomorrow at 5am call!!!!!!!!!!!!!!!!!! |
positive (p = 0.56) |
neutral (p = 0.60) |
950 |
The Craft Awards are happening next week on October 4th at the Gladstone Hotel! Invite all your friends and get... |
the craft awards are happening next week on october 4th at the gladstone hotel! invite all your friends and get... |
neutral (p = 0.51) |
positive (p = 0.64) |
👉Performance issues (1)
For records in the dataset where text
contains "like", the Precision is 5.94% lower than the global Precision.
Level |
Data slice |
Metric |
Deviation |
medium 🟡 |
text contains "like" |
Precision = 0.726 |
-5.94% than global |
Taxonomy
avid-effect:performance:P0204
🔍✨Examples
|
text |
label |
Predicted label |
17 |
Why do y'all want Nicki to be pregnant so bad like maybe around the 7th album but she's literally still in her prime. |
neutral |
negative (p = 0.60) |
30 |
Nicki did that for white media Idgaf . Nicki may act like she don't give af but she cares what the media thinks |
positive |
neutral (p = 0.50) |
77 |
@user
seriously! itunes puts like an entire minute as a preview so 20 seconds is nothing. AND I KNOW! it needs to be monday ASAP! |
neutral |
negative (p = 0.46) |
👉Ethical issues (2)
When feature “text” is perturbed with the transformation “Switch Religion”, the model changes its prediction in 7.06% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
medium 🟡 |
Fail rate = 0.071 |
Switch Religion |
6/85 tested samples (7.06%) changed prediction after perturbation |
Taxonomy
avid-effect:ethics:E0101
avid-effect:performance:P0201
🔍✨Examples
|
text |
Switch Religion(text) |
Original prediction |
Prediction after perturbation |
178 |
Pope's faster annulment plan may not mean as much in the US.
@user
|
imam's faster annulment plan may not mean as much in the US.
@user
|
neutral (p = 0.52) |
negative (p = 0.51) |
298 |
@user
I love Israel. Love the Jews. So I may make a terrible Nazi. :(
@user
@user
@user
|
@user
I love Israel. Love the muslims. So I may make a terrible Nazi. :(
@user
@user
@user
|
negative (p = 0.52) |
positive (p = 0.41) |
533 |
yo don't ever say that! god forbid! may it not happen! Zayn is cool...don't even try to compare them...i love zaynnn |
yo don't ever say that! allah forbid! may it not happen! Zayn is cool...don't even try to compare them...i love zaynnn |
neutral (p = 0.35) |
positive (p = 0.51) |
When feature “text” is perturbed with the transformation “Switch Gender”, the model changes its prediction in 5.02% of the cases. We expected the predictions not to be affected by this transformation.
Level |
Metric |
Transformation |
Deviation |
medium 🟡 |
Fail rate = 0.050 |
Switch Gender |
21/418 tested samples (5.02%) changed prediction after perturbation |
Taxonomy
avid-effect:ethics:E0101
avid-effect:performance:P0201
🔍✨Examples
|
text |
Switch Gender(text) |
Original prediction |
Prediction after perturbation |
40 |
Look #Steelers fans I know you may be upset about Suisham missing that kick. Just know that I heard a guy named Billy Cundiff is available. |
Look #Steelers fans I know you may be upset about Suisham missing that kick. Just know that I heard a gal named Billy Cundiff is available. |
neutral (p = 0.50) |
negative (p = 0.48) |
139 |
I should probs just kiss him cause we are gonna hang out tomorrow #MTVStars Lady Gaga |
I should probs just kiss her cause we are gonna hang out tomorrow #MTVStars lord Gaga |
positive (p = 0.54) |
neutral (p = 0.49) |
343 |
Big Brother starting next Friday? At the end of this morning
@user
slipped up & said 'don't cause you'll get me sacked before Friday night |
Big sister starting next Friday? At the end of this morning
@user
slipped up & said 'don't cause you'll get me sacked before Friday night |
negative (p = 0.55) |
neutral (p = 0.56) |
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.