Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Investigating the Multilingual Calibration Effects of Language Model Instruction-Tuning

Jerry Huang, Peng Lu, Qiuhao Zeng, Yusuke Iwasawa, Yutaka Matsuo, Sarath Chandar, Edison Marrese-Taylor, Irene Li

TL;DR

This work analyzes how instruction-tuning affects calibration of multilingual LLMs across 29 and 42 languages on the MMLU-ProX and GlobalMMLU benchmarks. It finds that instruction-tuning generally increases over-confidence and fails to improve accuracy in low-resource languages, leading to mis-calibration. The study demonstrates that label smoothing applied to the instruction-tuning data improves calibration with minimal accuracy trade-offs, offering a practical mitigation. Overall, the results highlight the importance of incorporating multilingual calibration considerations into training and evaluation to enhance reliability and fairness in downstream use.

Abstract

Ensuring that deep learning models are well-calibrated in terms of their predictive uncertainty is essential in maintaining their trustworthiness and reliability, yet despite increasing advances in foundation model research, the relationship between such large language models (LLMs) and their calibration remains an open area of research. In this work, we look at a critical gap in the calibration of LLMs within multilingual settings, in an attempt to better understand how the data scarcity can potentially lead to different calibration effects and how commonly used techniques can apply in these settings. Our analysis on two multilingual benchmarks, over 29 and 42 languages respectively, reveals that even in low-resource languages, model confidence can increase significantly after instruction-tuning on high-resource language SFT datasets. However, improvements in accuracy are marginal or non-existent, resulting in mis-calibration, highlighting a critical shortcoming of standard SFT for multilingual languages. Furthermore, we observe that the use of label smoothing to be a reasonable method alleviate this concern, again without any need for low-resource SFT data, maintaining better calibration across all languages. Overall, this highlights the importance of multilingual considerations for both training and tuning LLMs in order to improve their reliability and fairness in downstream use.

Investigating the Multilingual Calibration Effects of Language Model Instruction-Tuning

TL;DR

This work analyzes how instruction-tuning affects calibration of multilingual LLMs across 29 and 42 languages on the MMLU-ProX and GlobalMMLU benchmarks. It finds that instruction-tuning generally increases over-confidence and fails to improve accuracy in low-resource languages, leading to mis-calibration. The study demonstrates that label smoothing applied to the instruction-tuning data improves calibration with minimal accuracy trade-offs, offering a practical mitigation. Overall, the results highlight the importance of incorporating multilingual calibration considerations into training and evaluation to enhance reliability and fairness in downstream use.

Abstract

Ensuring that deep learning models are well-calibrated in terms of their predictive uncertainty is essential in maintaining their trustworthiness and reliability, yet despite increasing advances in foundation model research, the relationship between such large language models (LLMs) and their calibration remains an open area of research. In this work, we look at a critical gap in the calibration of LLMs within multilingual settings, in an attempt to better understand how the data scarcity can potentially lead to different calibration effects and how commonly used techniques can apply in these settings. Our analysis on two multilingual benchmarks, over 29 and 42 languages respectively, reveals that even in low-resource languages, model confidence can increase significantly after instruction-tuning on high-resource language SFT datasets. However, improvements in accuracy are marginal or non-existent, resulting in mis-calibration, highlighting a critical shortcoming of standard SFT for multilingual languages. Furthermore, we observe that the use of label smoothing to be a reasonable method alleviate this concern, again without any need for low-resource SFT data, maintaining better calibration across all languages. Overall, this highlights the importance of multilingual considerations for both training and tuning LLMs in order to improve their reliability and fairness in downstream use.
Paper Structure (32 sections, 1 theorem, 15 equations, 215 figures, 73 tables)

This paper contains 32 sections, 1 theorem, 15 equations, 215 figures, 73 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Uncertainty Calibration.
  4. Multilingual LLMs.
  5. Methodology
  6. Models and Setup.
  7. Tasks.
  8. Results
  9. Calibration Across Languages
  10. Accounting For Over-Confidence
  11. Label smoothing
  12. Can LS help with Multilingual Calibration?
  13. Discussion
  14. Why does SFT Lead to Mis-Calibration?
  15. On Difference Confidence Measures for LLMs.
  16. ...and 17 more sections

Key Result

Proposition B.3

A linear penalty (or a Lagrangian term) for the hard constraint $\bm{d}({\bm{x}}) = \bm{0}$ is bounded from above and below by $\mathrm{KL}\left({\bm{u}}\|\widehat{\bm{\sigma}}\left({\bm{x}};\bm{\theta}\right)\right)$, up to additive constants

Figures (215)

  • Figure 1: Comparison of Base models (red) and instruction-tuned (blue) models on various languages on GlobalMMLU. Deviation from the straight line indicates under or over-confidence; models show increasing overconfidence across all languages.
  • Figure 2: Reliability diagrams for the yo split of GlobalMMLU after instruction-tuning on Tulu3Mixture (top) and OpenHermes (bottom).
  • Figure 3: Reliability diagrams for the GlobalMMLU dataset for the am language.
  • Figure 4: Reliability diagrams for the GlobalMMLU dataset for the am language after instruction-tuning on the Tulu3Mixture dataset.
  • Figure 5: Reliability diagrams for the GlobalMMLU dataset for the am language after instruction-tuning on the OpenHermes dataset.
  • ...and 210 more figures

Theorems & Definitions (5)

  • Remark B.1
  • proof
  • Definition B.2
  • Proposition B.3
  • proof