toasterai/SambaInstruct-Alpha-v1-7B

SambaInstruct-Alpha

An experimental assistant model fine tuned with ZERO GPT-4/3.5/Claude/etc. data. Instead, data was obtained by using a causal model to complete turns. Something similar to URIAL.

The resulting dataset is more conversational and has a different feel from typical corporate model output. Right now, most of the resulting data is single-turn, is not of the highest quality and contains some hallucination-inducing examples. We are looking to release this data publicly under Apache 2.0 once these shortcomings are addressed.

Please use the ChatML template. The recommended prompt we trained is You are a helpful AI chatbot. and it probably will work best, but make sure to test different configurations.

The model might hallucinate some personhood (due to the data), but it should still refer to itself as an AI.

Look forward to future iterations of the model! Here is a list of future goals for them:

• Better base? Qwen3, Nemo, etc.
• Thinking mode/chain of thought.
• More specialized data (creative writing, roleplay, code, math etc.)
• More compliance and customizable alignment.