The LLMs have demonstrated exceptional performance on several tasks using an inference a few blows, also called learning in the context (ICL). The main problem is to select the most representative demonstrations from large training data sets. The first methods have selected demonstrations based on relevance using similarity scores between each example and the input question. Current methods suggest using additional selection rules, as well as similarity, to improve the effectiveness of demonstration selection. These improvements introduce significant general costs when the number of shots increases. The effectiveness of selected demonstrations should also take into account the specific LLM used, because different LLMs present variable capacities and areas of knowledge.
Researchers from the University of Shanghai Jiao Tong, Xiaohongshu Inc., Carnegie Mellon University, the University of Beijing, without affiliation, University College in London and the University of Bristol have proposed a dealer (little demonstration still essential predelector), a method for identifying a central subsense of demonstrations containing the most examples Representative in data training, adjusted to specific LLM. To build this subset, measures of “sufficiency” and “necessity” are introduced in the pre-selection stage, as well as an algorithm based on trees. In addition, the feeder reduces the size of the 20% training data while maintaining performance and integrating transparently with various demonstration selection techniques downstream in ICL through the LLM ranging from 300m to 8b parameters.
The eater is evaluated on 6 sets of text classification data: SST-2, SST-5, COLA, TREC, SUBJ and FPB, covering the tasks of the classification of feelings and linguistic analysis with the textual impact. It is also evaluated on the GSM8K reasoning data set, the SMCALFlow semantic time data set and the scientific data game of GPQA questions. Official divisions for each data set are directly followed to obtain training and test data. In addition, several LLM variants are used to assess the performance of the method, including two GPT-2, GPT-NEO variants with parameters of 1.3b, GPT-3 with 6B, Gemma-2 parameters with 2B, Llama-2 parameters with 7b settings, Llama-3 with LLM and Qwen-2.5 parameters with 32 parameters.
The results concerning learning performance in the context show that the feeder allows the retention of almost half of the training samples while achieving higher or comparable performance. The evaluation of performance with a few blows on complex tasks using LLM as Gemma-2 shows that the feeder improves performance even when LLMs fight with difficult tasks. It works effectively with a large number of shots, handling situations where LLM performance generally drops when the number of examples goes from 5 to 10 due to noisy or repeated demonstrations. In addition, the feeder minimizes the negative impact on LLM performance by evaluating the sufficiency and the need for each demonstration, and helps the LLM performance stability
When optimizing the organic level, the feeder performs improved performance using a high -quality data set for fine adjustment while simultaneously reducing calculation expenditure, aligning with the nucleus selection principle. The results indicate that the LLMS with fine adjustment provides greater performance improvements compared to the increase in LLMS with contexts, with power gains, with even better performance gains in fine settings. Performance analysis reveals that feeder's efficiency increases first, then decreases with an increasing number of races or towers (R and K, respectively), confirming that the identification of representative subsets of training data sets improves LLM performance. However, too narrow sub-assemblies can limit potential performance gains.
In conclusion, the researchers introduced Feeder, a predecessor demonstration designed to use LLM capacities and knowledge of the field to identify high -quality demonstrations thanks to an effective discovery approach. It reduces training data requirements while maintaining comparable performance, offering a practical solution for effective LLM deployment. Future research orientations include exploration of applications with larger LLMs and the extension of food capacities to areas such as data security and data management. Feeder makes a precious contribution to the selection of demonstration, offering researchers and practitioners an effective tool to optimize LLM performance while reducing the general calculation costs.
Discover the Paper. All the merit of this research goes to researchers in this project.
Meet the newsletter of AI dev read by 40K + developers and researchers from Nvidia, Openai, Deepmind, Meta, Microsoft, JP Morgan Chase, Amgen, Aflac, Wells Fargo and 100 others (Subscribe now)
Sajjad Ansari is a last year's first year of the Kharagpur Iit. As a technology enthusiast, he plunges into AI's practical applications by emphasizing the understanding of the impact of AI technologies and their real implications. It aims to articulate complex AI concepts in a clear and accessible way.