Good Robot 2 πŸ€–

The model "Good Robot" had one simple goal in mind: to be a good instruction-following model that doesn't talk like ChatGPT.

Built upon the Mistral 7b 0.2 base, this model aims to provide responses that are as human-like as possible, thanks to some DPO training using the (for now, private) minerva-ai/yes-robots-dpo dataset.

HuggingFaceH4/no-robots was used as the base for generating a custom dataset to create DPO pairs.

It should follow instructions and be generally as smart as a typical Mistral model - just not as soulless and full of GPT slop.

Changes from the original good-robot model:

  • Mistral 7b-0.2 base (32k native context, no SWA)
  • ChatML prompt format
  • Trained using GaLore method

Prompt Format:

ChatML

<|im_start|>system
System message
<|im_start|>user
User message<|im_end|>
<|im_start|>assistant

Credits:

Model made in collaboration with Gryphe.

Training Data:

Limitations:

While I did my best to minimize GPTisms, no model is perfect, and there may still be instances where the generated content has GPT's common phrases - I have a suspicion that's due to them being engrained into Mistral model itself.

License:

cc-by-nc-4.0

Downloads last month
16
Safetensors
Model size
7.24B params
Tensor type
FP16
Β·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Dataset used to train kubernetes-bad/good-robot-2