Confusion about Data Fields in v3

[FlowRays]

Hello

I’m currently working with the v3 version of the dataset and have a couple of questions regarding the data fields. I would greatly appreciate it if you could provide some clarification on the following points:

1. completion and response Fields:

   • I’m a bit confused about which of these fields represents the actual solution. Could you please explain the difference between them and how each is derived?

2. Mismatch in test and input_output Counts:

   • I noticed that the number of entries in the test field does not match the number in the input_output field. Could you kindly explain the reason for this discrepancy?

Thank you so much for your time and assistance! I look forward to your response.

Best regards

[FlowRays]

Another issue is that I noticed the LaTeX expressions in the problem descriptions are not being parsed correctly. For example:

• The description contains -109 <= nums[i] <= 109, but it’s clear that the intended meaning is -10^9 <= nums[i] <= 10^9.

This issue seems to be present across all LeetCode-related datasets on Hugging Face. If possible, I would be extremely grateful if this could be addressed. Fixing this would greatly improve the accuracy and usability of the datasets.

Thank you again for your time and effort in maintaining these datasets.

[FlowRays]

Also, I noticed that the question_id in the dataset goes up to 3481, which suggests it includes the latest LeetCode problems. However, some questions on the LeetCode website are missing in the dataset. Could you clarify why?

[FlowRays]

1. Based on my local testing, both completion and response are solutions.
2. In the test, for questions with question_id 349, 760, 970, 1169, and 1792, both completion and response resulted in AssertionError. I suspect that there might be an issue with these specific tests themselves.

[dangne]

Hi @FlowRays, thank you for the detailed analyses — these are very helpful. Regarding the incorrect constraints, I suspect that most LeetCode-related datasets use html2text to convert LeetCode problems from HTML to Markdown format, but they didn't enable the include_sup_sub setting, which led to incorrect conversions (e.g., LeetCodeHard-Gym is one example).

For example, the following code:

import html2text

h = html2text.HTML2Text()
# h.include_sup_sub is False by default
h.handle("<code>0 &lt;= grid[i][j] &lt;= 10<sup>5</sup></code>")

Will give the following output:

`0 <= grid[i][j] <= 105`\n\n

@newfacade Could you confirm this?

[yucc-leon]

Same questions as above.

Regarding LaTeX expressions that were not correctly parsed (mainly those involving scientific notation), I tried to correct them using GPT (considering OAI is expected to use properly formatted data and the O-series models are known for good error correction capabilities).
After some trials I found that O4-mini was relatively affordable and accurate (others might not find the errors or would rewrite the original description to different expressions). But it still required some post-processing.

And here is an alternative for the original dataset: https://huggingface.co/datasets/Leon-Leee/LeetCodeDataset_rectified
There are two additional columns: problem_typo_check (Yes/No, but not always correct, so ignore it), problem_description_corrected (the rewritten texts).
One may consider using problem_description_corrected instead of problem_description as the former corrects some of the errors by GPT.
However, I haven't checked each entry of the results so there might be unexpected new errors. Please let me know if you find any of them.