search

SkywardAI / kirin

Self-hosted and local-first application for inference and RAG on consumer grade hardware.
Apache License 2.0
7 stars 8 forks source link

base-line LLM #108

Closed Aisuko closed 5 months ago

Aisuko commented 5 months ago

Minimal size, and RLHF pre-trained model which are with labels 'trl', 'text-generation-inference'

Aisuko commented 5 months ago

Currently, we use get-2 as the base model. As we discussed, we will migrate to the model which focus more on dialogue. And we are interested in huge giant LLM with many layers redundancy.

So, I think we can implement a base line evaluate system to support us complete this task in a reasonable way. And the little bit out of date it ok for our situation.

The candidates can be

Aisuko commented 5 months ago

We will create a smaller model by using transfer learning techniques on the base-line model

Aisuko commented 5 months ago

https://www.kaggle.com/code/aisuko/goal-directed-dialog-godel

Aisuko commented 5 months ago

https://github.com/SkywardAI/kimchima/pull/75