sharegpt format support

[IvanDeng0]

Before Asking 在提问之前

• I have read the README carefully. 我已经仔细阅读了 README 上的操作指引。

• I have pulled the latest code of main branch to run again and the problem still existed. 我已经拉取了主分支上最新的代码，重新运行之后，问题仍不能解决。

Search before asking 先搜索，再提问

• I have searched the Data-Juicer issues and found no similar questions. 我已经在 issue列表 中搜索但是没有发现类似的问题。

Question

my dataset is formed with sharegpt format, just like:
[
{"conversations": [
{"from": "user", "value": "..."},
{"from": "gpt", "value": "..."},
]
}
]

how to set the "text-keys"

Additional 额外信息

No response

[HYLcool]

Hi @IvanDeng0 , thanks for your suggestion!

As far as I know, the format of ShareGPT is the same as that of LLaVA. If so, we already support the llava-to-dj and dj-to-llava format conversion tools. You can find it in the LLaVA-like format section from this document and check if it can meet your needs. If not, contact us and discuss your needs with us here.

If there are needs on SFT dataset format, please follow up. @drcege

[lostkevin]

For complex use, sharegpt-format dataset can contain multiple images or videos like DJ format while LLaVA-format seems not. A more extensive script may be required for conversion. Reference: https://github.com/hiyouga/LLaMA-Factory/tree/main/data#multimodal-image-dataset-1

[yxdyc]

Thank you for your questions and suggestions! We have previously engaged in detailed design and discussions regarding the compatibility of the DJ schema with Modelscope-Swift and LLaMA-Factory, with the goal of supporting a wide range of post-tuning datasets and convenient training by ensuring compatibility with these two libraries.

An initial development effort has been made under PR #454 , where alignment for Modelscope-Swift's query/response format has been implemented, which should be easily to extended for other formats like Alpaca Format (refer to the AutoPreprocessor). @HYLcool and I are working on more ready-to-use bidirectional conversion tools, firstly the Sharegpt Format.

Welcome any discussions or PRs regarding this issue and next steps :)

Cc @cyruszhang: this might be useful for our AutoFormatter

[yxdyc]

Update: Initially implemented for sharegpt-format in #514, with a preliminary format specification. We are refactoring and will provide DJ's DatasetLoaders and AutoFormatters in the near future, in collaboration with @cyruszhang .