@inproceedings{shum-etal-2024-first,
title = "{FIRST}: Teach A Reliable Large Language Model Through Efficient Trustworthy Distillation",
author = "Shum, KaShun and
Xu, Minrui and
Zhang, Jianshu and
Chen, Zixin and
Diao, Shizhe and
Dong, Hanze and
Zhang, Jipeng and
Raza, Muhammad Omer",
editor = "Al-Onaizan, Yaser and
Bansal, Mohit and
Chen, Yun-Nung",
booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2024",
address = "Miami, Florida, USA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.emnlp-main.703",
doi = "10.18653/v1/2024.emnlp-main.703",
pages = "12646--12659",
abstract = "Large language models (LLMs) have become increasingly prevalent in our daily lives, leading to an expectation for LLMs to be trustworthy {---}- both accurate and well-calibrated (the prediction confidence should align with its ground truth correctness likelihood). Nowadays, fine-tuning has become the most popular method for adapting a model to practical usage by significantly increasing accuracy on downstream tasks. Despite the great accuracy it achieves, we found fine-tuning is still far away from satisfactory trustworthiness due to {``}tuning-induced mis-calibration{''}. In this paper, we delve deeply into why and how mis-calibration exists in fine-tuned models, and how distillation can alleviate the issue. Then we further propose a brand new method named Efficient Trustworthy Distillation (FIRST), which utilizes a small portion of teacher{'}s knowledge to obtain a reliable language model in a cost-efficient way. Specifically, we identify the {``}concentrated knowledge{''} phenomenon during distillation, which can significantly reduce the computational burden. Then we apply a {``}trustworthy maximization{''} process to optimize the utilization of this small portion of concentrated knowledge before transferring it to the student. Experimental results demonstrate the effectiveness of our method, where better accuracy (+2.3{\%}) and less mis-calibration (-10{\%}) are achieved on average across both in-domain and out-of-domain scenarios, indicating better trustworthiness.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shum-etal-2024-first">
<titleInfo>
<title>FIRST: Teach A Reliable Large Language Model Through Efficient Trustworthy Distillation</title>
</titleInfo>
<name type="personal">
<namePart type="given">KaShun</namePart>
<namePart type="family">Shum</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minrui</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianshu</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zixin</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shizhe</namePart>
<namePart type="family">Diao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hanze</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jipeng</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="given">Omer</namePart>
<namePart type="family">Raza</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yaser</namePart>
<namePart type="family">Al-Onaizan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohit</namePart>
<namePart type="family">Bansal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yun-Nung</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Miami, Florida, USA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Large language models (LLMs) have become increasingly prevalent in our daily lives, leading to an expectation for LLMs to be trustworthy —- both accurate and well-calibrated (the prediction confidence should align with its ground truth correctness likelihood). Nowadays, fine-tuning has become the most popular method for adapting a model to practical usage by significantly increasing accuracy on downstream tasks. Despite the great accuracy it achieves, we found fine-tuning is still far away from satisfactory trustworthiness due to “tuning-induced mis-calibration”. In this paper, we delve deeply into why and how mis-calibration exists in fine-tuned models, and how distillation can alleviate the issue. Then we further propose a brand new method named Efficient Trustworthy Distillation (FIRST), which utilizes a small portion of teacher’s knowledge to obtain a reliable language model in a cost-efficient way. Specifically, we identify the “concentrated knowledge” phenomenon during distillation, which can significantly reduce the computational burden. Then we apply a “trustworthy maximization” process to optimize the utilization of this small portion of concentrated knowledge before transferring it to the student. Experimental results demonstrate the effectiveness of our method, where better accuracy (+2.3%) and less mis-calibration (-10%) are achieved on average across both in-domain and out-of-domain scenarios, indicating better trustworthiness.</abstract>
<identifier type="citekey">shum-etal-2024-first</identifier>
<identifier type="doi">10.18653/v1/2024.emnlp-main.703</identifier>
<location>
<url>https://aclanthology.org/2024.emnlp-main.703</url>
</location>
<part>
<date>2024-11</date>
<extent unit="page">
<start>12646</start>
<end>12659</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T FIRST: Teach A Reliable Large Language Model Through Efficient Trustworthy Distillation
%A Shum, KaShun
%A Xu, Minrui
%A Zhang, Jianshu
%A Chen, Zixin
%A Diao, Shizhe
%A Dong, Hanze
%A Zhang, Jipeng
%A Raza, Muhammad Omer
%Y Al-Onaizan, Yaser
%Y Bansal, Mohit
%Y Chen, Yun-Nung
%S Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing
%D 2024
%8 November
%I Association for Computational Linguistics
%C Miami, Florida, USA
%F shum-etal-2024-first
%X Large language models (LLMs) have become increasingly prevalent in our daily lives, leading to an expectation for LLMs to be trustworthy —- both accurate and well-calibrated (the prediction confidence should align with its ground truth correctness likelihood). Nowadays, fine-tuning has become the most popular method for adapting a model to practical usage by significantly increasing accuracy on downstream tasks. Despite the great accuracy it achieves, we found fine-tuning is still far away from satisfactory trustworthiness due to “tuning-induced mis-calibration”. In this paper, we delve deeply into why and how mis-calibration exists in fine-tuned models, and how distillation can alleviate the issue. Then we further propose a brand new method named Efficient Trustworthy Distillation (FIRST), which utilizes a small portion of teacher’s knowledge to obtain a reliable language model in a cost-efficient way. Specifically, we identify the “concentrated knowledge” phenomenon during distillation, which can significantly reduce the computational burden. Then we apply a “trustworthy maximization” process to optimize the utilization of this small portion of concentrated knowledge before transferring it to the student. Experimental results demonstrate the effectiveness of our method, where better accuracy (+2.3%) and less mis-calibration (-10%) are achieved on average across both in-domain and out-of-domain scenarios, indicating better trustworthiness.
%R 10.18653/v1/2024.emnlp-main.703
%U https://aclanthology.org/2024.emnlp-main.703
%U https://doi.org/10.18653/v1/2024.emnlp-main.703
%P 12646-12659
Markdown (Informal)
[FIRST: Teach A Reliable Large Language Model Through Efficient Trustworthy Distillation](https://aclanthology.org/2024.emnlp-main.703) (Shum et al., EMNLP 2024)
ACL
- KaShun Shum, Minrui Xu, Jianshu Zhang, Zixin Chen, Shizhe Diao, Hanze Dong, Jipeng Zhang, and Muhammad Omer Raza. 2024. FIRST: Teach A Reliable Large Language Model Through Efficient Trustworthy Distillation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 12646–12659, Miami, Florida, USA. Association for Computational Linguistics.