Question about mistralai/Mistral-7B-Instruct-v0.1 and its base model
Dear [Developer/Team],
I recently came across mistralai/Mistral-7B-Instruct-v0.1, and it has been very helpful in my project.
To ensure proper usage, I want to clarify its relationship with mistralai/Mistral-7B-v0.1:
Direct Fine-tuning: Is mistralai/Mistral-7B-Instruct-v0.1 a direct fine-tuned version of mistralai/Mistral-7B-v0.1, or were there intermediate models/checkpoints involved?
Inheritance: Does it keep the same architecture and weights as mistralai/Mistral-7B-v0.1?
This will help me build on it correctly.
I really appreciate your help.
Hey, (not part of the team) but to answer your questions, It's likely they had intermediate checkpoints during SFT (internal to the team), but the official/final instruct checkpoint is this one. It's a instruct fine-tuned version of the base Mistral-7B-v0.1
Yes it keeps the same architecture as Mistral-7B-v0.1, you can even check from the config.json or in the transformers repo.
Weights are not identical since it has been SFT'ed, but they are compatible (whether base or instruct, both can be loaded) because it's the same architecture.
Hope that helps, @juliendenize or @pandora-s could confirm if needed.