Quantization made by Richard Erkhov.
The model is a work in progress it is inteneded to be used as a base model to enable the automation of agents that need to work in an accounting environment where work is provided in the form of complex problems and decisions need to be made for the purpose of posting into a GL given the double entry system as well as IFRS. I did not train this model up on GAAP, I might for a future use case do so.
This model is highly experimental and is most definately not to be considered production ready. I will throw more resources at this model as time goes by.
In my view it does pretty well.
When prompting the model add in the context of the model "How must this be treated from accounting perspective:
"I used a self curated chain of thought so it would be good to in the comments on the model indicate where the model might not be technically correct in its responses or where it went off script, same as "logging a ticket", I will look at the instances and enhance the training set so that the model becomes more accurate and usefull over time.
Please remember this model is part of an experiment and can not be considered production ready, use at your own risk if you decide to deploy in environment where output would be used to inform decisions.
This qwen2 model was trained 2x faster with
Unsloth and Huggingface's TRL library.
