Keyword "assistant" Error with ShareGPT Datasets

[MilesQLi]

⚠️ Please check that this feature request hasn't been suggested before.

• I searched previous Ideas in Discussions didn't find any similar feature requests.
• I searched previous Issues didn't find any similar feature requests.

🔖 Feature description

On ShareGPT datasets such as shibing624/sharegpt_gpt4, some "from" values are "assistant" rather than "gpt", so, the program raises the keyword error.

✔️ Solution

Maybe take both assistant and gpt as the machine's responses?

❓ Alternatives

No response

📝 Additional Context

No response

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this feature has not been requested yet.
• I have provided enough information for the maintainers to understand and evaluate this request.

[NanoCode012]

I checked that linked dataset. It seems to be human/gpt.

In case you want to swap keys, please check this.

https://github.com/OpenAccess-AI-Collective/axolotl/blob/21cf09b60840ee03ba3ee4e57c2707d2296f532c/src/axolotl/prompt_strategies/sharegpt.py#L78-L90

It swaps roles using a simple dictionary map.

[MilesQLi]

@NanoCode012 I said "some "from" values are "assistant" rather than "gpt"". You checked only some data from that dataset, not all.

[NanoCode012]

Oh I see! You can use this kind of role_map then.

role_map = {"assistant": "gpt", "human": "human", "gpt": "gpt"} 

[MilesQLi]

@NanoCode012 Thanks! I'm not sure I'm doing it the way you mentioned. I updated the role_map, but still get the error. Please see the two screenshots.

[MilesQLi]

I fixed this by updating SimpleShareGPTPromptTokenizingStrategy with the role_map you provided.