We constantly need to keep up with this ever-changing world, as do machine learning models, to produce precise output. Large Language Models often suffer from fallacy issues; that is, they are unaware of unseen events or generate text with incorrect facts owing to the outdated/noisy data. For example- LLMs such as ChatGPT and LlaMA possess information only up to their last training point, and essentially, we need to update the parametric knowledge within the LLMs to modify their specific behaviors. Numerous knowledge editing or model editing methods have been introduced to craft edits in machine learning models whilst minimizing the impact on unrelated inputs.
To tackle persistent challenges based on knowledge cut-off/biased outputs, researchers have applied two major methods:
- Fine – Tuning, traditional fine-tuning, and delta tuning utilize domain-specific datasets, but they consume enormous resources and risk the potential of catastrophic forgetting.
- Prompt- Augmentation, when provided with ample demonstrations or gathered contexts, large language models (LLMs) exhibit the capacity to improve their reasoning capabilities and enhance their generation tasks through integrating external knowledge. The downside is this technique may be sensitive to factors such as the prompting template and the selection of in-context examples.
Owing to significant differences among various knowledge editing methods and the variations in task setups, no standard implementation framework is available. To address these issues and provide a unified framework, researchers have introduced EASYEDIT, an easy-to-use knowledge editing framework for LLMs. It supports cutting-edge knowledge editing approaches and can be readily applied to many well-known LLMs such as T5, GPT-J, and LlaMA.
The EASYEDIT platform introduces a user-friendly “edit” interface that enables easy model modification. Comprising key elements like Hparams, Method, and Evaluate, this interface seamlessly incorporates various strategies for knowledge editing. The core mechanism for implementing these strategies is the APPLY_TO_MODEL function, accessible through different methods. The figure above demonstrates an instance of applying MEND to LLaMA, altering the output of the U.S. President to Joe Biden.
EASYEDIT employs a modular approach to organizing editing methods and evaluating their efficacy while also accounting for their interplay and combination. The platform accommodates a range of editing scenarios, encompassing single-instance, batch-instance, and sequential editing. Furthermore, it conducts evaluations of critical metrics such as Reliability, Generalization, Locality, and Portability, which assist users in identifying the most suitable method tailored to their distinct requirements.
The knowledge editing results on LlaMA-2 with EASYEDIT demonstrate that knowledge editing surpasses traditional fine-tuning regarding reliability and generalization. In conclusion, the EasyEdit framework emerges as a pivotal advancement in the realm of large language models (LLMs), addressing the critical need for accessible and intuitive knowledge editing.
Check out the Paper and Github. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 29k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.
If you like our work, please follow us on Twitter
Janhavi Lande, is an Engineering Physics graduate from IIT Guwahati, class of 2023. She is an upcoming data scientist and has been working in the world of ml/ai research for the past two years. She is most fascinated by this ever changing world and its constant demand of humans to keep up with it. In her pastime she enjoys traveling, reading and writing poems.
