Table tick ✅ and cross❌ symbols

API
1

for table-based questions OpenAI gpt3.5 turbo Api finds it difficult to answer questions.

e.g.: if a table chunk contains tick (:white_check_mark:) or cross (:x:) symbol, OpenAI is not able to answer such questions.

input is in html format, with the table tags.

2

Please provide a complete example of an input which results in the problem you are experiencing.

3

Well, if it was this forum’s HTML, it wouldn’t just be regex to replace the unicode with another character or “True” that the AI can understand, it would be replacing <img src="https://emoji.discourse-cdn.com/twitter/white_check_mark.png?v=12" title=":white_check_mark:" class="emoji" alt=":white_check_mark:" loading="lazy" width="20" height="20" style="aspect-ratio: 20 / 20;">…

4

Sure, but it’s there any reason to suspect it’s not just the actual symbol?

1000016514
10000165141080×2222 319 KB

That’s why I’m asking for a reproducible example from the OP because, from the information provided, it’s not clear what the best solution would be.

It’s also not even clear if these are the exact symbols present in their table or if they are convenient stand-ins given the resources of the forum software.

1 Like
5

hi @all, thanks for your support,

e.g.:

image
image961×169 7.43 KB

question: is automation supported in iOS.
answer: yes, it is supported.

I can’t use Unicode- since they come under emoji tag.

question: is automation supported in iOS. Define the symbol.
answer: yes, it is supported, the symbols provided is ’ :x: ’

In my understanding, OpenAI gpt3.5 is not able to understand the symbols.
well.

6

What I’m looking for from you here is an exact example of a text input you are giving to the model that is failing for you.

Please copy the exact HTML, and between two sets of triple backticks, share it here.

Like this,

```
Your EXACT HTML text goes here... 
```

That way there is no possibility for confusion about what you are sending or why you’re encountering issues.

7

here in the input format. @anon22939549

<table><tr></tr><tr><td> </td><td>Desktop/Browser</td><td>iOS</td><td>Android</td></tr><tr><td>Automations</td><td>✅</td><td>❌ </td><td>❌ </td></tr><tr><td>Interface Designer</td><td>✅</td><td>❌ </td><td>❌ </td></tr></table>

<table>
  <tr></tr>
  <tr>
    <td></td>
    <td>Desktop/Browser</td>
    <td>iOS</td>
    <td>Android</td>
  </tr>
  <tr>
    <td>Automations</td>
    <td>✅</td>
    <td>❌ </td>
    <td>❌ </td>
  </tr>
  <tr>
    <td>Interface Designer</td>
    <td>✅</td>
    <td>❌ </td>
    <td>❌ </td>
  </tr>
</table>
8

I edited your post to include a nicely formatted version to make it easier for others to read.

I have found that the models tend to struggle a bit more with HTML tables than they do with \LaTeX or Markdown tables.

Here is the equivalent Markdown table,

|                    | Desktop/Browser | iOS | Android |
|--------------------|-----------------|-----|---------|
| Automations        | ✅              | ❌  | ❌      |
| Interface Designer | ✅              | ❌  | ❌      |

That said, I tested 3.5 on ChatGPT and in the API with your exact input and it nailed the questions every time.

I tested the API in the playground with no system message at several different temperatures. It’s possible whatever system message you’re using is leading it astray, or it could have simply been a bad roll of the dice.

If you asked all the questions in the same context, the previous responses would have impacted the following responses.

I guess I would suggest you run a few more tests, try converting the table to markdown if the issue persists, then if it still isn’t resolved come back with all the additional information you were able to collect.

(Come back anyway if it resolves itself to let us know, and hopefully stick around!)

2 Likes
9

thanks @anon22939549 , will sure check it.