Fix KeyError occurring using fine_tunes.prepare_data

[serinamarie]

Description

KeyError occurs when running openai tools fine_tunes.prepare_data -f training_file.jsonl when training_file.jsonl contains a prompt/completion that is BOTH a duplicate and a long example. Without trying to change too much, this fix would:

• wrap the retrieval of long_examples and long_indexes into a function, calling it preemptively within long_examples_validator to provide analysis information about how many rows are long examples, then calling it when actually dropping rows within optional_fn and providing info to the user if the keys that are being dropped have changed.

Related Issue

Fixes #121

Other Notes

I am also happy to provide a file that can be used to reproduce this error.

Example Output:

When the error would normally occur, instead you would see:

❯ openai tools fine_tunes.prepare_data -f training_file_0927.jsonl
Analyzing...
- There are 2 duplicated prompt-completion sets. These are rows: [5, 6]
- There are 2 examples that are very long. These are rows: [3, 6]
Based on the analysis we will perform the following actions:
- [Recommended] Remove 2 duplicate rows [Y/n]: y
- [Recommended] Remove 2 long examples [Y/n]: y
The indices of the long examples has changed as a result of a previously applied recommendation.
The 1 long examples to be dropped are now at the following indices: [3]
- [Recommended] Add a suffix separator ` ->` to all prompts [Y/n]: y 

[hallacy]

Awesome! Sorry for the delay. A few nits and then I think this is good to go. Can you explain a bit more about why this works?

[serinamarie]

Awesome! Sorry for the delay. A few nits and then I think this is good to go. Can you explain a bit more about why this works?

Hi @hallacy, it works because in the original code when you are presented with the indices to drop for long examples (originally at line 164), at that point in time no recommendations have been applied. For example, a row that is both a long example AND a duplicate may exist at index 4; and the current script, when removing the duplicates, will remove index 4, but index 4 is still 'scheduled' to be deleted when dropping long examples at original line 170. When the optional_fn() is applied, the key no longer exists, causing a KeyError.
tl;dr sometimes optional_fn() will try to delete an index that doesn't exist anymore because the long_indexes have been calculated before any recommendations have been applied.

By recalculating the long_indexes when actually applying the optional_fn() on the df, we ensure that we have an up-to-date array containing the indices where each long example exists, rather than a possibly outdated one. I hope that helps to clear it up; I could have included that in the description :)

[hallacy]

Love it! Thank you!