Investigate natural langugage processing shcemes private LLM development using LLMs. Our aims are to explore the state-of-the-art natural language processing algorithms using LLMs for achieving effective long-term memory and private systems(LLM/RAG) and improving performance as in:
1. Motivation
- needs for long-term memory
2. Research Goal and Issue
- Goal : Develop a mechanism for long-term memory
- Issues
3. Approach
- Develop LLM long-term memory
4. Result
- Effectiveness : applying model editing to reflect dialogue history into weights of distilGPT2, we found we were able to successfully get it to answer with facts it previously did not know
- Overfitting issue : after model editing distilGPT2 with MEND we found significant overfitting such that it would answer with the edited response to many unrelated questions
- General dialogue performance : the quality ofthe dialogue generated by distilGPT2 is quite low, we found GPT2-large to generate much more coherent dialogue so we tested our model editing memory mechanism applied to GPT2-large and results are better but still room for improvement
- Multi-edit functionality required : our current prototype only enables single edits, but to refelct the facts from the entire dialogue history we need to enable multiple edits. Performance evaluation with memory benchmarks also required multi-edit funcionality
