@inproceedings{ahrabian-etal-2024-adaptation,
title = "On the Adaptation of Unlimiformer for Decoder-Only Transformers",
author = "Ahrabian, Kian and
Benhaim, Alon and
Patra, Barun and
Pujara, Jay and
Singhal, Saksham and
Song, Xia",
editor = "Calzolari, Nicoletta and
Kan, Min-Yen and
Hoste, Veronique and
Lenci, Alessandro and
Sakti, Sakriani and
Xue, Nianwen",
booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
month = may,
year = "2024",
address = "Torino, Italia",
publisher = "ELRA and ICCL",
url = "https://aclanthology.org/2024.lrec-main.1085",
pages = "12395--12402",
abstract = "One of the prominent issues stifling the current generation of large language models is their limited context length. Recent proprietary models such as GPT-4 and Claude 2 have introduced longer context lengths, 8k/32k and 100k, respectively; however, despite the efforts in the community, most common models, such as LLama-2, have a context length of 4k or less. Unlimiformer (Bertsch et al., 2023) is a recently popular vector-retrieval augmentation method that offloads cross-attention computations to a kNN index. However, its main limitation is incompatibility with decoder-only transformers out of the box. In this work, we explore practical considerations of adapting Unlimiformer to decoder-only transformers and introduce a series of modifications to overcome this limitation. Moreover, we expand the original experimental setup on summarization to include a new task (i.e., free-form Q{\&}A) and an instruction-tuned model (i.e., a custom 6.7B GPT model). Our results showcase the effectiveness of these modifications on summarization, performing on par with a model with 2x the context length. Moreover, we discuss limitations and future directions for free-form Q{\&}A and instruction-tuned models.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ahrabian-etal-2024-adaptation">
<titleInfo>
<title>On the Adaptation of Unlimiformer for Decoder-Only Transformers</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kian</namePart>
<namePart type="family">Ahrabian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alon</namePart>
<namePart type="family">Benhaim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Barun</namePart>
<namePart type="family">Patra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jay</namePart>
<namePart type="family">Pujara</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Saksham</namePart>
<namePart type="family">Singhal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xia</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicoletta</namePart>
<namePart type="family">Calzolari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Min-Yen</namePart>
<namePart type="family">Kan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Veronique</namePart>
<namePart type="family">Hoste</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">Lenci</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakriani</namePart>
<namePart type="family">Sakti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nianwen</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA and ICCL</publisher>
<place>
<placeTerm type="text">Torino, Italia</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>One of the prominent issues stifling the current generation of large language models is their limited context length. Recent proprietary models such as GPT-4 and Claude 2 have introduced longer context lengths, 8k/32k and 100k, respectively; however, despite the efforts in the community, most common models, such as LLama-2, have a context length of 4k or less. Unlimiformer (Bertsch et al., 2023) is a recently popular vector-retrieval augmentation method that offloads cross-attention computations to a kNN index. However, its main limitation is incompatibility with decoder-only transformers out of the box. In this work, we explore practical considerations of adapting Unlimiformer to decoder-only transformers and introduce a series of modifications to overcome this limitation. Moreover, we expand the original experimental setup on summarization to include a new task (i.e., free-form Q&A) and an instruction-tuned model (i.e., a custom 6.7B GPT model). Our results showcase the effectiveness of these modifications on summarization, performing on par with a model with 2x the context length. Moreover, we discuss limitations and future directions for free-form Q&A and instruction-tuned models.</abstract>
<identifier type="citekey">ahrabian-etal-2024-adaptation</identifier>
<location>
<url>https://aclanthology.org/2024.lrec-main.1085</url>
</location>
<part>
<date>2024-05</date>
<extent unit="page">
<start>12395</start>
<end>12402</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T On the Adaptation of Unlimiformer for Decoder-Only Transformers
%A Ahrabian, Kian
%A Benhaim, Alon
%A Patra, Barun
%A Pujara, Jay
%A Singhal, Saksham
%A Song, Xia
%Y Calzolari, Nicoletta
%Y Kan, Min-Yen
%Y Hoste, Veronique
%Y Lenci, Alessandro
%Y Sakti, Sakriani
%Y Xue, Nianwen
%S Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)
%D 2024
%8 May
%I ELRA and ICCL
%C Torino, Italia
%F ahrabian-etal-2024-adaptation
%X One of the prominent issues stifling the current generation of large language models is their limited context length. Recent proprietary models such as GPT-4 and Claude 2 have introduced longer context lengths, 8k/32k and 100k, respectively; however, despite the efforts in the community, most common models, such as LLama-2, have a context length of 4k or less. Unlimiformer (Bertsch et al., 2023) is a recently popular vector-retrieval augmentation method that offloads cross-attention computations to a kNN index. However, its main limitation is incompatibility with decoder-only transformers out of the box. In this work, we explore practical considerations of adapting Unlimiformer to decoder-only transformers and introduce a series of modifications to overcome this limitation. Moreover, we expand the original experimental setup on summarization to include a new task (i.e., free-form Q&A) and an instruction-tuned model (i.e., a custom 6.7B GPT model). Our results showcase the effectiveness of these modifications on summarization, performing on par with a model with 2x the context length. Moreover, we discuss limitations and future directions for free-form Q&A and instruction-tuned models.
%U https://aclanthology.org/2024.lrec-main.1085
%P 12395-12402
Markdown (Informal)
[On the Adaptation of Unlimiformer for Decoder-Only Transformers](https://aclanthology.org/2024.lrec-main.1085) (Ahrabian et al., LREC-COLING 2024)
ACL
- Kian Ahrabian, Alon Benhaim, Barun Patra, Jay Pujara, Saksham Singhal, and Xia Song. 2024. On the Adaptation of Unlimiformer for Decoder-Only Transformers. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12395–12402, Torino, Italia. ELRA and ICCL.