tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
MV The maxoverseparate optimal Vshaped scores applied to explicit rubrics
AQ The average scoring rule over singledimensional quadrati c scoring rules on each explicit rubric
We ﬁrst compare AQ and MV scoring rules on numerical feedback The correlation is calculated
between the algorithmic score for numerical review and the i nstructor score for numerical review
We check the expost fairness in score in Table 6 Our results show that while being optimal for
incentivizing binary eﬀort MV does not guarantee expost fa irness in score We observe that AQ
outperforms MV in simulating TA preference for both algorit hm classes While MV scores the
most surprising dimension AQ compares all dimensions The correlation between instructor score
and quadratic score shows the instructor is combining the di stance on all dimensions and scoring
similarly to quadratic scoring rule
As a benchmark for textual scores we use AQ for comparison
23B Prompts
Appendix B1 presents the detailed prompts that we use in Eli citationGPT Appendix B2 shows
the prompts for direct GPT queries of scores
B1 Prompts in ElicitationGPT
In this section we present the implementation details of El icitationGPT When querying ChatGPT
we follow prompt engineering tricks including chain of tho ught and task splitting OpenAI 2023
We include the queries and answer formats for GPT here To avo id random outputs we set
parameter toppto 0001 for temperature
The system is initialized with a system prompt
System Prompt You are helping summarizing points in reviews of student hom ework
For each instructor review Ii query GPT with a summarization task as a warm up An example
query is the following
Prompt Review 1
AnswerAlgorithm
The answers are all correct and clearly presented
ProofAnalysis
For part 1 a little more elaboration is needed for the ﬁnal st atement
For part 2 the claim then fngn is not correct There i s a big gap from there to the
ﬁnal result
Clarity
The solution is clearly presented
Response Pros
All answers are correct
The solutions are clearly presented
Cons
More elaboration is needed for the ﬁnal statement in part 1
The claim fngn in part 2 is incorrect leading to a gap in the ﬁnal result
After summarizing all instructor reviews on the same homewo rk assignment GPT is queried to
summarize all the instructor reviews into a list of topics W e query it twice with an opportunity
to improve the summarization with all previous messages ap pended in history
Prompt Generating Topics
The reviews above are for 8 diﬀerent submissions on the same ho mework assignment Now
your task is to help cluster the points with similar meanings
24Your output should be formatted as the following 1 include a numbered list of points 2
each point is followed by the reviews which mention them whe re R1 means review 1 3 After
each review mark if this is a negative opinion or a positive o pinion N for negative P for
positive Separate PN from review number Rs with a comma Ou tput as fewer clusters as
possible
For example your output should be formatted as
1 AnswerAlgorithm
R1 P The answers are all correct and clearly presented
R2 N The answer is wrong
2 Clarity
R3 N However there is some typo in part 1 that mixes with par t 2 which makes the
solution a little confusing
Response 1 AnswerAlgorithm
R1 P The answers are all correct and clearly presented
R2 P The answers are all correct and clearly presented
R3 P The answers are all correct and clearly presented th ough there is a typo in part
1
R4 P The answers are correct
R5 N For part 1 According to the rubric you get 0 by claimi ng an incorrect result
as correct and give a proof for it
R6 P You answers are perfectly correct
R7 P Part A is correct
R8 P The answers are all correct and clearly presented
2 ProofAnalysis
R1 N For part 1 a little more elaboration is needed for the ﬁnal statement For part
2 the claim then fngn is not correct
R2 N In part a theres an issue in the reasoning fn an d gn are both On does
not imply fn Ogn
R3 N The claim f 2fnwere O2gn then limit equal 0 i s incorrect
R4 N Proof by example is not suﬃce R5 N The claim given fn Ogn as
n gnfn is incorrect
R6 P The proof and analysis is correct
R7 N The proof for part A is not suﬃcient You need to prove t here does not exist c
and n0
R8 P The proof is nice and succinct
3 Clarity
R1 P The solution is clearly presented
R2 N It would be better if the proof and analysis are better structured so the line of
reasoning for each claim is more clear
25 R3 N The submission is acceptable overall However ther e is some typo in part 1
that mixes with part 2 which make the solution a little confus ing
R4 P The submission is clear overall despite the issue
R5 P The overall clarity is good R6 N Please clearly an d explicitly state the
TrueFalse conclusion before starting present your proof
R7 P The clarity for part A is good
R8 N The clarity of part a could be better The clarity of pa rt b is perfect
Prompt Revise topics
Are there rubric points with similar meanings that can be com bined Revise your answer
and combine rubrics points if any Output in the same format a s above
Response
In the example conversation above the answer to the summari zation prompt here is the same
as the explicit rubrics Generally when the homework submi ssion includes code plot etc the
summarization may also include these implicit topics like c ode quality plot quality etc
Now we query GPT to generate indicator states Take AnswerA lgorithm as example We
also query it twice with an opportunity to improve the summar ization with all previous messages
appended in history The trick works well

MV The maxoverseparate optimal Vshaped scores applied to explicit rubrics

AQ The average scoring rule over singledimensional quadrati c scoring rules on each explicit rubric

We ﬁrst compare AQ and MV scoring rules on numerical feedback The correlation is calculated

between the algorithmic score for numerical review and the i nstructor score for numerical review

We check the expost fairness in score in Table 6 Our results show that while being optimal for

incentivizing binary eﬀort MV does not guarantee expost fa irness in score We observe that AQ

outperforms MV in simulating TA preference for both algorit hm classes While MV scores the

most surprising dimension AQ compares all dimensions The correlation between instructor score

and quadratic score shows the instructor is combining the di stance on all dimensions and scoring

similarly to quadratic scoring rule

As a benchmark for textual scores we use AQ for comparison

23B Prompts

Appendix B1 presents the detailed prompts that we use in Eli citationGPT Appendix B2 shows

the prompts for direct GPT queries of scores

B1 Prompts in ElicitationGPT

In this section we present the implementation details of El icitationGPT When querying ChatGPT

we follow prompt engineering tricks including chain of tho ught and task splitting OpenAI 2023

We include the queries and answer formats for GPT here To avo id random outputs we set

parameter toppto 0001 for temperature

The system is initialized with a system prompt

System Prompt You are helping summarizing points in reviews of student hom ework

For each instructor review Ii query GPT with a summarization task as a warm up An example

query is the following

Prompt Review 1

AnswerAlgorithm

The answers are all correct and clearly presented

ProofAnalysis

For part 1 a little more elaboration is needed for the ﬁnal st atement

For part 2 the claim then fngn is not correct There i s a big gap from there to the

ﬁnal result

Clarity

The solution is clearly presented

Response Pros

All answers are correct

The solutions are clearly presented

Cons

More elaboration is needed for the ﬁnal statement in part 1

The claim fngn in part 2 is incorrect leading to a gap in the ﬁnal result

After summarizing all instructor reviews on the same homewo rk assignment GPT is queried to

summarize all the instructor reviews into a list of topics W e query it twice with an opportunity

to improve the summarization with all previous messages ap pended in history

Prompt Generating Topics

The reviews above are for 8 diﬀerent submissions on the same ho mework assignment Now

your task is to help cluster the points with similar meanings

24Your output should be formatted as the following 1 include a numbered list of points 2

each point is followed by the reviews which mention them whe re R1 means review 1 3 After

each review mark if this is a negative opinion or a positive o pinion N for negative P for

positive Separate PN from review number Rs with a comma Ou tput as fewer clusters as

possible

For example your output should be formatted as

1 AnswerAlgorithm

R1 P The answers are all correct and clearly presented

R2 N The answer is wrong

2 Clarity

R3 N However there is some typo in part 1 that mixes with par t 2 which makes the

solution a little confusing

Response 1 AnswerAlgorithm

R1 P The answers are all correct and clearly presented

R2 P The answers are all correct and clearly presented

R3 P The answers are all correct and clearly presented th ough there is a typo in part

1

R4 P The answers are correct

R5 N For part 1 According to the rubric you get 0 by claimi ng an incorrect result

as correct and give a proof for it

R6 P You answers are perfectly correct

R7 P Part A is correct

R8 P The answers are all correct and clearly presented

2 ProofAnalysis

R1 N For part 1 a little more elaboration is needed for the ﬁnal statement For part

2 the claim then fngn is not correct

R2 N In part a theres an issue in the reasoning fn an d gn are both On does

not imply fn Ogn

R3 N The claim f 2fnwere O2gn then limit equal 0 i s incorrect

R4 N Proof by example is not suﬃce R5 N The claim given fn Ogn as

n gnfn is incorrect

R6 P The proof and analysis is correct

R7 N The proof for part A is not suﬃcient You need to prove t here does not exist c

and n0

R8 P The proof is nice and succinct

3 Clarity

R1 P The solution is clearly presented

R2 N It would be better if the proof and analysis are better structured so the line of

reasoning for each claim is more clear

25 R3 N The submission is acceptable overall However ther e is some typo in part 1

that mixes with part 2 which make the solution a little confus ing

R4 P The submission is clear overall despite the issue

R5 P The overall clarity is good R6 N Please clearly an d explicitly state the

TrueFalse conclusion before starting present your proof

R7 P The clarity for part A is good

R8 N The clarity of part a could be better The clarity of pa rt b is perfect

Prompt Revise topics

Are there rubric points with similar meanings that can be com bined Revise your answer

and combine rubrics points if any Output in the same format a s above

Response

In the example conversation above the answer to the summari zation prompt here is the same

as the explicit rubrics Generally when the homework submi ssion includes code plot etc the

summarization may also include these implicit topics like c ode quality plot quality etc

Now we query GPT to generate indicator states Take AnswerA lgorithm as example We

also query it twice with an opportunity to improve the summar ization with all previous messages

appended in history The trick works well