TeG-Instruct: Towards Premium Instruction-Tuning Data via Text-Grounded Task Design

Yongrui Chen, Haiyun Jiang, Xinting Huang, Shuming Shi, Guilin Qi

Primary Area: datasets and benchmarks

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Natural Language Processing, Large Language Model, Instruction Tuning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

The enhancement of language model capabilities heavily relies on the availability of high-quality instruction-tuning data. However, current data collection approaches face limitations due to the high costs associated with manual labeling or the hallucination of relying solely on LLMs.
To overcome these challenges, this paper proposes a scalable solution for automatically gathering top-notch instruction-tuning data. The method involves training LLMs to autonomously design tasks utilizing human-written texts, thereby aiding LLMs in mitigating erroneous outputs.
In contrast to other approaches that utilize human-written texts, our method employs a task generator capable of simultaneously producing the instruction, input, and output.
It aims to minimize the introduction of noise from the original text and ensure coherent and aligned task components.
Additionally, we train a discriminator to identify and filter out invalid tasks that might contain misleading or hallucination, thus further improving the quality of the collected data.
The results of the automated and manual evaluation experiments validate the reliability and validity of our proposed dataset, demonstrating the applicability it brings to LLMs for both in-domain and out-of-domain generation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5018