Word Search (0.0075 acc, 0.30 f1)

[azhx]

Thank you for contributing an eval! ♥️

🚨 Please make sure your PR follows these guidelines, failure to follow the guidelines below will result in the PR being closed automatically. Note that even if the criteria are met, that does not guarantee the PR will be merged nor GPT-4 access granted. 🚨

PLEASE READ THIS:

In order for a PR to be merged, it must fail on GPT-4. We are aware that right now, users do not have access, so you will not be able to tell if the eval fails or not. Please run your eval with GPT-3.5-Turbo, but keep in mind as we run the eval, if GPT-4 gets higher than 90% on the eval, we will likely reject since GPT-4 is already capable of completing the task.

We plan to roll out a way for users submitting evals to see the eval performance on GPT-4 soon. Stay tuned! Until then, you will not be able to see the eval performance on GPT-4. We encourage partial PR's with ~5-10 example that we can then run the evals on and share the results with you so you know how your eval does with GPT-4 before writing all 100 examples.

Eval details 📑

Eval name

word-search

Eval description

This evaluation tests whether GPT can solve a word search given in text format. There are 4 types of prompts included in the dataset of 400 examples (100 each).

• difficulty 1: 1-shot word search with words given
• difficulty 2: 1-shot word search with words not given
• difficulty 3: 0-shot word search with words given
• difficulty 4: 0-shot word search with words not given

Below is an example of a difficulty 3 word search.

---------------------------
        WORD SEARCH        
---------------------------
G M M N E H H Z S P H N W X
N H Y Z W I N G Y Z N J X U
D H P E S O J D T D K X X Q
J E G D N I X Z E K G Q Q B
W G N M V N T W D E O D T G
L D W S K C W R M S Z C S L
N Z N S H R W C K B M S A G
B C G G L E U D V J B L P E
B S N A L A B P C I B P D R
A T J U C S Z O X O C K F U
C X D G R I O C Y X N S P E
A Z F E I N I A W A Y C F F
P U X G H G F Y R K Z F F A
N D A U D N R W P W L E B X

Find these words: GAUGE, INCREASING, JOSEPH, PAST, RANK
* Words can go W, NE, S, NW, SW, SE, N, and, E.

Answer according to the following format, where coordinates are in the form (1-indexed col # from left, 1-indexed row # from top):
Answer Key: DANCE SE @ (6, 9), IPOD S @ (2, 5), JUMP NW @ (8, 8), PEACEFUL N @ (12, 13), STUDIES S @ (13, 3)
Answer Key:

What makes this a useful eval?

This task requires many layers of reasoning abilities and from gpt-3.5 tests has proven to be very difficult to achieve. Over a couple tests, Gpt-3.5 has been able to achieve (0.0075 acc, 0.30 f1) at best

Criteria for a good eval ✅

Below are some of the criteria we look for in a good eval. In general, we are seeking cases where the model does not do a good job despite being capable of generating a good response (note that there are some things large language models cannot do, so those would not make good evals).

Your eval should be:

• [x] Thematically consistent: The eval should be thematically consistent. We'd like to see a number of prompts all demonstrating some particular failure mode. For example, we can create an eval on cases where the model fails to reason about the physical world.
• [x] Contains failures where a human can do the task, but either GPT-4 or GPT-3.5-Turbo could not.
• [x] Includes good signal around what is the right behavior. This means either a correct answer for Basic evals or the Fact Model-graded eval, or an exhaustive rubric for evaluating answers for the Criteria Model-graded eval.
• [x] Include at least 100 high quality examples (it is okay to only contribute 5-10 meaningful examples and have us test them with GPT-4 before adding all 100)

If there is anything else that makes your eval worth including, please document it below.

Unique eval value

Not only does this eval test the ability for GPT to locate specific words in a word search (pattern identification and reasoning), it also tests its ability to reason about whether some sequence of letters is any notable word in the case where the words are not given. This is a simple, menial task for humans given sufficient time as we simply need to check each of 8 directions at every location and see if the word being read out is substantial and makes sense

Eval structure 🏗️

Your eval should

• [x] Check that your data is in evals/registry/data/{name}
• [x] Check that your yaml is registered at evals/registry/evals/{name}.jsonl
• [x] Ensure you have the right to use the data you submit via this eval

(For now, we will only be approving evals that use one of the existing eval classes. You may still write custom eval classes for your own cases, and we may consider merging them in the future.)

Final checklist 👀

Submission agreement

By contributing to Evals, you are agreeing to make your evaluation logic and data under the same MIT license as this repository. You must have adequate rights to upload any data used in an Eval. OpenAI reserves the right to use this data in future service improvements to our product. Contributions to OpenAI Evals will be subject to our usual Usage Policies (https://platform.openai.com/docs/usage-policies).

• [x] I agree that my submission will be made available under an MIT license and complies with OpenAI's usage policies.

Email address validation

If your submission is accepted, we will be granting GPT-4 access to a limited number of contributors. Access will be given to the email address associated with the merged pull request.

• [x] I acknowledge that GPT-4 access will only be granted, if applicable, to the email address used for my merged pull request.

Limited availability acknowledgement

We know that you might be excited to contribute to OpenAI's mission, help improve our models, and gain access to GPT-4. However, due to the requirements mentioned above and high volume of submissions, we will not be able to accept all submissions and thus not grant everyone who opens a PR GPT-4 access. We know this is disappointing, but we hope to set the right expectation before you open this PR.

• [x] I understand that opening a PR, even if it meets the requirements above, does not guarantee the PR will be merged nor GPT-4 access granted.

Submit eval

• [x] I have filled out all required fields in the evals PR form
• [x] (Ignore if not submitting code) I have run pip install pre-commit; pre-commit install and have verified that black, isort, and autoflake are running when I commit and push

Failure to fill out all required fields will result in the PR being closed.

Eval JSON data

Since we are using Git LFS, we are asking eval submitters to add in as many Eval Samples (at least 5) from their contribution here:

View evals in JSON

Eval

{"input": [{"role": "system", "content": "You are a helpful assistant. Follow the user's examples and instructions to solve problems."}, {"role": "user", "content": "---------------------------\n        WORD SEARCH        \n---------------------------\nX X U Q Q X Z I B F E C K M\nK C I H C Y U B J U E K O S\nO E U V Y U M U X D G H W U\nL V D G R A I G N M X I N B\nD T N P M D M S U X B Q L U\nP L W C Y J P Q U O N A S E\nW T S K A O R R E M P H L O\nM H M K R A E F J O M U Q I\nK R L L H J S L B V M A P L\nP Q A D B I S X L U F N R N\nE R N L Z L I T H H H M R Y\nS M G F H H V D Q P T L E B\nZ R I A P N E E K T Z H Z S\nH H S C B O X Q Q W C K T X\n\nFind these words: BALI, CHICK, IMPRESSIVE, SIGNAL, SUMMARY\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key: BALI SE @ (11, 5), CHICK W @ (5, 2), IMPRESSIVE S @ (7, 4), SIGNAL N @ (3, 14), SUMMARY SE @ (8, 5)\n-----------------------------\n         WORD SEARCH         \n-----------------------------\nF I G N P R N G Z V Q J B C M\nT X I X D Q B P K V D I C V Q\nB G A D S W A E T U O J K U N\nZ F A I N N I A Q V D U I W W\nG E R S R E G O R Y L E X K B\nT Q J Z F S O D L I D N Y G A\nA U R Y G Y W R S N Y J T N D\nA N T I C I P A T E D H Y I N\nK K W I K I B U M N U C M L V\nE W C A Q C M W L Q K T B E O\nX X C M T Y K E S H E D G V I\nQ X U A S L B C I S Q I G A B\nC L Q N T F A A S K Z Y P R N\nM U N X I I H L N Q R F W T U\nV Y W F X T E F Y X R G L U E\n\nFind these words: ANTICIPATED, DILDOS, LACE, ROGERS, TRAVELING\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key:"}], "ideal": ["ANTICIPATED E @ (1, 8)", "DILDOS W @ (11, 6)", "LACE N @ (8, 14)", "ROGERS W @ (9, 5)", "TRAVELING N @ (14, 14)"]}
{"input": [{"role": "system", "content": "You are a helpful assistant. Follow the user's examples and instructions to solve problems."}, {"role": "user", "content": "---------------------------\n        WORD SEARCH        \n---------------------------\nQ S R P Q J M B H R J Z P Q\nB T M H X S O N Z R O B Z W\nL C Y M K L X I D T A V F A\nF M M S P A L E S T I N E X\nQ O K P U T A Y M R E P S U\nU U R F O F V Z W S C L U T\nX E D T G Q I A B V J A U H\nZ A B V Y X G X U I C I H Z\nQ F P L H V S Y W Z Q N P L\nG R S C J X D D N J Z O Y J\nW F Y U U Y V M N R P L U U\nQ F F Z R L D B V U A O S G\nS P U M X Y T V N Q Z C R L\nK J L R U C Z T L G P T N G\n\nFind these words: COLONIAL, CULT, FORTY, PALESTINE, SPERM\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key: COLONIAL N @ (12, 13), CULT SE @ (4, 10), FORTY SE @ (1, 4), PALESTINE E @ (5, 4), SPERM W @ (13, 5)\n-------------------------------\n          WORD SEARCH          \n-------------------------------\nV M Z C F P Z Z N W U O U L J Q\nL I F A F V W R S F Y W R L L D\nR O K K I J W X P H N S I H W Y\nJ A L S A I L I N G G E C E P B\nK R N N O N J M I A D O D M F R\nV Q C Y C U H K R A N D L O J Y\nD A Q H M N N E V T B R T R A C\nX D K C N O J O R E I E B K E F\nY P F N J C R I D V E E N U L Q\nV C E E P U B E I M N E R I T Q\nN H C R V U Y I R X I O A X X D\nD V Y F T C Q I H H R B B L G S\nC J B I W A E S F T I H Y U U P\nZ X N D B Q L J B P L Z J O H P\nY G L C U Y Q A N Q V R Z I E C\nH E Y I Y A L K K X U V L S U K\n\nFind these words: ANYMORE, CONTRIBUTING, FRENCH, SAILING, TIRE\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key:"}], "ideal": ["ANYMORE SE @ (2, 4)", "CONTRIBUTING SW @ (13, 4)", "FRENCH N @ (4, 12)", "SAILING E @ (4, 4)", "TIRE W @ (15, 10)"]}
{"input": [{"role": "system", "content": "You are a helpful assistant. Follow the user's examples and instructions to solve problems."}, {"role": "user", "content": "---------------------------\n        WORD SEARCH        \n---------------------------\nX F D X U J F W Z F A C C J\nI P V J W I G M K G S X Y Y\nB J F S I M O Q D X H F A N\nR A R R E S T E D S J F L E\nI A A K E G Y E I X Q G B H\nT I W R H T S W J Y G E R Z\nA J C J U S L C B H D M S P\nN A N V E R M T Z D F F T J\nN O K L Q O Y B O O U E C Q\nI Q B R I Y J F W T H C I U\nC B Y G J O G F U Z G Q V S\nA B G E A S C R J H C N R G\nU F W A Y Q E J D I L V U P\nK J U R J Q A Z A P P Y W Q\n\nFind these words: ACRES, ARRESTED, BLESSED, BRITANNICA, FUTURE\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key: ACRES NE @ (2, 8), ARRESTED E @ (2, 4), BLESSED NE @ (3, 10), BRITANNICA S @ (1, 3), FUTURE SW @ (12, 8)\n---------------------------\n        WORD SEARCH        \n---------------------------\nU E S P I F U Q X O A X S J\nQ O L T L Q G D K Q B X J Q\nD L A I Q A A Q O Z S W J V\nX T U C O S N C E F O M O L\nT W D N Q W Y O V N L R K G\nR E K E A U G W S N U Y C W\nK P I N T H B B X R T F J D\nR V D I D S Z F U E E D I P\nT F O U H X U Q A K E P M B\nZ C O B A P J J W F F E J L\nB S W Y U R U N D Y Z C V T\nW J J K W L I E G A J Z N W\nM V O Q Z S Z A W W H F O E\nI C A Z O R P G O Y L N B U\n\nFind these words: ABSOLUTE, ADJUSTED, JEFF, PERSONAL, PROZAC\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key:"}], "ideal": ["ABSOLUTE S @ (11, 1)", "ADJUSTED NW @ (10, 12)", "JEFF W @ (13, 10)", "PERSONAL NW @ (12, 9)", "PROZAC W @ (7, 14)"]}
{"input": [{"role": "system", "content": "You are a helpful assistant. Follow the user's examples and instructions to solve problems."}, {"role": "user", "content": "---------------------------\n        WORD SEARCH        \n---------------------------\nW R H N H R P L W R N O G Y\nG L M D D B E G R N U K H H\nK X S C E M Z T Y H S P Y A\nN L Q Y A T D R S Y J P D W\nP V X N G U A V Q A G G R R\nI T W T Q W U C H C F K O O\nE Q G M H A K P I Y E E G C\nR E J O W X A I G D C D E L\nR E W Y T E L O C F E N N A\nE Y P R O H J F V K O D Z Q\nE W M P Z N Q X I O Q S W Z\nX C B T A Q U U O E D B B W\nL S D J I I M C S I Y C V N\nB P Q I A Q J U S O V C I M\n\nFind these words: COLE, DEDICATED, FASTER, HYDROGEN, PIERRE\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key: COLE W @ (9, 9), DEDICATED NW @ (12, 10), FASTER NW @ (11, 6), HYDROGEN S @ (13, 2), PIERRE S @ (1, 5)\n-----------------------------\n         WORD SEARCH         \n-----------------------------\nP B I T J Y I G X S D G X X A\nP C K M P S C A N N E R C T O\nN B W D T J U C H R L C I X N\nN H V I L P M X H D T Z V Y I\nF N T Y A G R I C U L T U R E\nR O P D R S Z Y P Y A U T K N\nE S X Y E E A S A M H T S A O\nC W T K A N R U D F B X P Z P\nS N A F L Y A U O G M O I C T\nQ S B E O X U T T Q J T P S B\nY A O V C B N L S A R X L U P\nS P P I P V M I I W E Q Q S L\nB M G J E U Y N S V K F K S Y\nI B K Y N N M K G Y K I V E V\nN B H Y S B I D Y K D T G X E\n\nFind these words: AGRICULTURE, ASTHMA, FEATURE, SCANNER, SUSSEX\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key:"}], "ideal": ["AGRICULTURE E @ (5, 5)", "ASTHMA W @ (14, 7)", "FEATURE NW @ (12, 13)", "SCANNER E @ (6, 2)", "SUSSEX S @ (14, 10)"]}
{"input": [{"role": "system", "content": "You are a helpful assistant. Follow the user's examples and instructions to solve problems."}, {"role": "user", "content": "---------------------------\n        WORD SEARCH        \n---------------------------\nB L R H K S R C S Q H O F O\nB G P B X V J V F A A H D S\nG C J L U H R G Y Y S B N K\nM Y G Z I P R O W L I Y K N\nY J D O R A T L J L S E J D\nS S F A P B U K K A W F V B\nK M Y H X Z U E S I Y J H T\nA P I H R H C Q F C S F O P\nB C B H B D L H M E W C N B\nS P D J I M A J N P Q R K U\nS R J S L V D T Q S Q D F F\nV I D Z Y X M V K E J Q W Q\nP I E T K C E E N R V P F A\nN P H E N O M E N O N N T V\n\nFind these words: ESPECIALLY, GRAPHICS, PHENOMENON, PRAY, SENT\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key: ESPECIALLY N @ (10, 12), GRAPHICS SW @ (8, 3), PHENOMENON E @ (2, 14), PRAY SW @ (6, 4), SENT SW @ (11, 8)\n---------------------------\n        WORD SEARCH        \n---------------------------\nE E C P L V B Q K S Q H L P\nP C A R Q N G V E H E E S E\nI G C P I L U K O G O H M D\nJ L T O Z N T S Z A M M D U\nM R I R L D H O I H C T D L\nA U Z T P O F I Z Q H N L C\nY T L S O H N F P L I A I X\nV X V S O D K R K Q I T A E\nY P J E U B J V K B W L U P\nK L K N S I K F L R D U X H\nL X B G C C C C Z U M S R Q\nL R B C F C D I N S C N K Z\nT F Q T V T S M K H P O R V\nL O F O G E D P X J W C R T\n\nFind these words: BRUSH, COLON, CONSULTANT, EXCLUDE, PORTS\n* Words can go W, NE, S, NW, SW, SE, N, and, E.\n\nAnswer Key:"}], "ideal": ["BRUSH S @ (10, 9)", "COLON SE @ (3, 3)", "CONSULTANT N @ (12, 14)", "EXCLUDE N @ (14, 8)", "PORTS S @ (4, 3)"]}

[azhx]

The script I'm contributing relies on an external library to generate the word search, so unsure of how to proceed.

[usama-openai]

Thank you for opening this PR. We're not accepting evals that have custom code implementations at this moment (but we are accepting custom model-graded evals).

If possible, could you rewrite this eval using an existing template (like Match, FuzzyMatch, or Includes) or using ModelGraded logic?

Thanks

Closing this PR, please open another PR with the suggested changes.

[azhx]

Hi, doesn't this already use FuzzyMatch? I even tested it using the testing scripts

[Ein-Tim]

@azhx

The problem is the following (quote @usama-openai):

We're not accepting evals that have custom code implementations at this moment (but we are accepting custom model-graded evals).

[usama-openai]

Closing the PR due to inactivity; please reopen if you get a chance to address comments.