Pith. sign in

REVIEW 4 major objections 7 minor 37 references

The Breeze 2 Herd of Models: Traditional Chinese LLMs Based on Llama with Vision-Aware and Function-Calling Capabilities

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims Breeze2 3B and 8B are the strongest Traditional Chinese function-calling and image-understanding models in their size class, with the 8B close to GPT-4o-mini.

desk verdict The Breeze2 models are a real, useful release for Traditional Chinese NLP, but the abstract's 'strongest in its size class' claim is not backed by the paper's own tables. read the letter →

arxiv 2501.13921 v3 pith:X4UIW47E submitted 2025-01-23 cs.CL

MediaTek Research: Chan-Jan Hsu , Chia-Sheng Liu , Meng-Hsi Chen , Muxi Chen , Po-Chun Hsu , Yi-Chang Chen , Da-Shan Shiu This is my paper · ORCID
classification cs.CL
keywords TraditionalChineseLLMfunctioncallingvision-languagemodelLlamamultimodalinstructiontuningTaiwanknowledgedecisiontokenmobiledeployment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Breeze2 is a two-model suite, with 3B and 8B parameters, built by taking Llama 3.2 3B and Llama 3.1 8B and continuing pre-training on about 900 GB of Traditional Chinese text before adding vision and function-calling modules. The paper's central claim is that, without reasoning-inducing prompts, these are the strongest models in their size class for Traditional Chinese function calling and image understanding at the time of publication. The evidence is a set of benchmarks covering Taiwan general knowledge, instruction following, long context, function calling in English and Traditional Chinese, and vision-language understanding, plus case studies and a mobile app demo. If the claim holds, developers get a publicly released small model that approaches commercial API performance on tool use and understands Taiwan-specific language and images better than larger open rivals.

What carries the argument

The load-bearing mechanism is the staged training recipe rather than a single formula: continued pre-training on a Taiwan-focused corpus, two-phase vision alignment (first only the MLP projector, then full weights), and a combined post-training pass over text instruction, visual instruction, and function-calling data. The named mechanism on the function-calling side is the Decision Token: with functions supplied, the model's first token is either <|use_tool|> or <|answer|>, and this choice also lets the authors create negative examples by removing the relevant function from a call example. Function-calling post-training data is built from 110k English instances, 19k Traditional Chinese translations, and 10k non-call examples generated by removing the relevant function from call examples. On the vision side, the InternViT-300M-448px encoder was chosen for resolution and for its stronger recognition of Chinese characters in images, and it is connected to the Llama backbone by an MLP projector.

What would settle it

An independent rerun that pits the released Breeze2 checkpoints, GPT-4o-mini, and every other openly available 3B-8B model against a newly constructed Traditional Chinese function-calling suite and a released TMMBench would settle the claim: if any comparable open model outranks Breeze2 on either, or if Breeze2 8B drops below GPT-4o-mini on the ZHTW leaderboard, the central claim fails.

Watch

Extended reading notes

Core claim

Starting from the Llama 3.2 3B and Llama 3.1 8B backbones, the authors continued pre-training on a curated Taiwan-focused corpus, then aligned an InternViT-300M-448px vision encoder to the language model through an MLP projector in two phases, and finally post-trained the combined model on text instruction, visual instruction, and function-calling data. For function calling, the model first emits a Decision Token, either <|use_tool|> or <|answer|>, so it can decide whether any supplied function applies before generating a call; this token is also used to manufacture non-function-call examples from function-call data. On the reported benchmarks, Breeze2 8B scores 87 overall accuracy on the BFCL benchmark versus GPT-4o-mini's 87, and 80 on the Traditional Chinese ZHTW leaderboard versus 81, while Breeze2 8B beats the larger Llama-3.2 11B on MMMU and on the paper's unreleased TMMBench. The paper's conclusion is that Breeze2 models are the strongest in their size class for Traditional Chinese function calling and image understanding at publication time.

Load-bearing premise

The headline comparison depends on the fairness and completeness of the two Traditional Chinese measurement tools, the ZHTW function-calling leaderboard and TMMBench, both built by the same team that trained the models, with TMMBench not yet released, so if those instruments omit strong competitors or favor the authors' training recipe, the 'strongest in its size class' conclusion would not follow even though the models may still be useful.

Editorial extensions

If this is right

  • If the central claim is correct, developers can build Traditional Chinese virtual assistants with close-to-commercial function calling and competitive image understanding using a publicly released 8B model.
  • The 3B variant is small enough to run on a phone with CPU inference, so on-device Traditional Chinese assistants are feasible, though the paper reports on-device throughput only for the language-only version.
  • The decision-token design makes relevance detection an explicit first step, which should reduce the common failure mode of hallucinating a function call when no supplied function fits.
  • The reported TMMLU+ gains over the base Llama models imply that continued pre-training on a curated 900 GB Traditional Chinese corpus transfers substantial general knowledge, not just lexical fluency.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same continued-pretraining-plus-decision-token recipe could plausibly transfer to other under-resourced languages with a similarly curated corpus and a script-aware vision encoder, but the paper does not test that transfer.
  • The paper's mobile demo stops at the language-only 3B model, so the untested case of on-device vision-language inference is the natural next measurement; the released checkpoints make it directly testable.
  • The decision-token mechanism could extend to multi-step agentic loops, choosing between answering, calling one tool, or calling several tools in parallel, though the paper only evaluates single-turn and parallel-call settings.
  • Releasing TMMBench, as the authors promise, would allow outsiders to check whether the Taiwan-specific vision gains are robust to independent construction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper introduces Breeze2, a suite of 3B and 8B multimodal language models built on Llama 3.2/3.1 bases, continued pre-trained on about 900 GB of Traditional Chinese text, and then adapted for vision understanding and function calling. The authors evaluate the models on Taiwan general knowledge (TMMLU+), instruction following (MT-Bench-tw), long-context passkey retrieval, function calling (BFCL and a ZHTW translation of BFCL), and vision benchmarks (MMMU and a newly proposed TMMBench). The abstract claims that, absent reasoning-inducing prompts, Breeze2 are the strongest models in Traditional Chinese function calling and image understanding in their size class. The models, the mobile application, and some benchmark infrastructure are released.

Significance. If the headline claim were established, Breeze2 would be a valuable open-source contribution for Traditional Chinese NLP and on-device deployment. The paper is transparent about training data composition, compute budgets, and architecture; it releases models and code; and it reports benchmark numbers across several tasks. However, as presented, the evidence does not yet support the 'strongest in its size class' claim: the baseline sets are too sparse to define the size class, and the key Traditional-Chinese-specific measurement instrument (TMMBench) is unreleased and constructed by the same team. The practical models may be genuinely useful, but the scientific case for the headline requires additional comparisons and independent evaluation.

major comments (4)
  1. [Abstract; §6.3; §6.4; Tables 7 and 8] The claim that Breeze2 are 'the strongest performing models in Traditional Chinese function calling and image understanding in its size class' is not grounded in the reported comparisons. Table 7 provides only one open 7B-class comparator (Gorilla-v2 7B) and no 3B-class models for ZHTW function calling; Table 8 contains no 3B- or 8B-class vision-language models, with Llama-3.2 11B as the nearest open model. The paper gives no definition of 'size class' or a systematic enumeration of candidate models, so the claim's scope is unverifiable. Notably, in Table 7, GPT-4o-mini outscores Breeze2 8B on overall accuracy (81 vs 80), and in Table 8, Llama-3.2 11B beats Breeze2 3B on both MMMU (41.7 vs 38.2) and TMMBench (42.4 vs 41.7). The conclusion's more cautious phrasing ('comparable to GPT-4o-mini', 'outperforms the larger Llama-3.2 11B') is consistent with these numbers but not with the abstract. Please either expand the baseline set to include the relevant open 3B/8B multimodal and function-calling models and formally define the size class, or revise the abstract to a weaker but supportable claim.
  2. [§6.4; Table 8] The Traditional Chinese vision benchmark TMMBench is constructed by the same team and is not released ('We plan to release TMMBench in the near future'); the GitHub link in footnote 9 appears to be a placeholder. Because the 'image understanding' half of the headline claim rests on TMMBench, the result is currently not independently reproducible or auditable. Please release TMMBench and its evaluation harness prior to the final version, or otherwise present vision evidence on an already-public benchmark with Traditional Chinese content.
  3. [§6.3; §5.4] The Function Calling Leaderboard for ZHTW (Lee et al., 2024) is also from MediaTek Research, and the function-calling training data (FC-TC-19k, FC-NF-10k) were generated using the synthetic translation method from the same group's Breeze-FC paper (Chen et al., 2024b). This overlap creates a risk that the evaluation and training share the same formatting and annotation distribution, potentially inflating Breeze2's scores relative to models trained with other formats. The paper should explicitly discuss this risk and, ideally, add an external evaluation (e.g., human scoring or an independently constructed ZHTW function-calling test) to break the circularity concern.
  4. [Abstract; §6.3] The abstract qualifies the headline claim with 'absent reasoning-inducing prompts', but this term is not defined or operationalized anywhere in the paper. It is unclear what prompting condition was used for all models and whether the same condition was applied to the baselines. Please define the term and state the exact prompting protocol used in the evaluations.
minor comments (7)
  1. [§1] The section title 'Indroduction' contains a typo; it should be 'Introduction'.
  2. [§6] The phrase 'genera knowledge' should be 'general knowledge'.
  3. [Author Contributions] The word 'cabaility' should be 'capability'.
  4. [§2] The architecture description says Breeze2 is built on 'Llama 3' but then specifies Llama 3.2 3B and Llama 3.1 8B; please clarify the relationship to the Llama 3 family and why the two sizes use different base versions.
  5. [Table 4] GPT-4o-mini's MMLU score (42.62) is much lower than typically reported and lower than Breeze2 3B's 59.9; please specify the exact evaluation setting (e.g., 0-shot, prompt language) and verify that it is comparable to the other models' settings.
  6. [§5.1] The decision to place function descriptions in the system role with a Decision Token would benefit from a brief summary of the evidence from Chen et al. (2024b) or an ablation; the current text cites the paper but does not give the key numbers.
  7. [Table 9] The mobile app's vision-aware performance is deferred to future work; if the paper claims a multimodal mobile experience, at least preliminary VLM latency and memory numbers should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the model's performance claims are empirical evaluations, and the same-team benchmarks, while a validity concern, are not equivalent by construction to the training inputs.

full rationale

Breeze2 is a model development report rather than a derivation chain, so the circularity patterns of interest (self-definitional identities, fitted parameters renamed as predictions, uniqueness theorems imported by self-citation) do not apply. The headline 'strongest in its size class' claim rests on benchmark tables (Tables 7 and 8), not on an equation that reduces to training data. The ZHTW function-calling leaderboard (Lee et al., 2024) and TMMBench (Section 6.4) are same-team measurement instruments, and TMMBench is explicitly unreleased ('We plan to release TMMBench in the near future'), which is a legitimate external-validity concern, but the paper does not state or imply that the benchmark questions were constructed from the training labels or that the scores are fitted outputs. The method borrows prior MediaTek components (Breeze-FC prompt templates, Decision Token, synthetic translation, Chen et al., 2024b), but citing one's own earlier method is not circular unless the cited result is itself the claim being proved; here the function-calling capability is independently evaluated on translated BFCL tasks. Similarly, the vision models are trained on public and in-house data and then tested on MMMU and TMMBench; no equation or fitting step makes the test score equal to a training objective. The sparse same-size baselines and unreleased benchmark weaken the strength of the comparative conclusion, but under the hard rules that concern is a correctness/evidence issue, not circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 1 invented entities

The paper is an empirical model release with no equations. The load-bearing assumptions are the validity of the evaluation benchmarks (some self-built and unreleased), the sufficiency of the baseline set, and the effectiveness of the data mixture. A single hand-chosen hyperparameter set is the main free parameter, reported without ablation.

free parameters (1)
  • Training hyperparameters = LR 1e-5 (text pre-training), 1e-6 (vision alignment and post-training); global batch 1024; 4 epochs post-training
    Chosen by hand without reported ablations (Sections 4.4 and 5.5). The claim 'strongest in size class' is not tested for sensitivity to these choices, so they are unexamined degrees of freedom in the recipe.
assumptions (3)
  • domain assumption The self-built benchmarks (Function Calling Leaderboard for ZHTW and TMMBench) are valid and unbiased measures of Traditional Chinese function calling and vision understanding.
    Section 6.3 uses the ZHTW leaderboard from Lee et al. (2024), whose author list overlaps with this paper (Yi-Chang Chen and Da-Shan Shiu). Section 6.4 uses TMMBench, which the same team constructed and has not released. If these benchmarks are biased toward the authors' own training recipes, the headline claim is not established.
  • domain assumption The comparison set (Llama 3.2 3B/11B, Llama 3.1 8B, Breeze 7B, Gorilla-v2 7B, GPT-4o-mini) is sufficient to establish 'strongest in its size class'.
    Section 6 reports results only against these baselines; other multilingual or multimodal models of 3B/8B scale (for example Qwen2.5-VL, InternVL2, or MiniCPM-V) are absent, so the size-class claim is untested against the most likely competitors.
  • domain assumption Continued pretraining on the described 900GB corpus improves Traditional Chinese capabilities without degrading English abilities unacceptably.
    Section 4 assumes the corpus mixture and training recipe produce the observed improvements; no ablation or controlled comparison with alternative corpora is provided.
invented entities (1)
  • TMMBench
    purpose: A Traditional Chinese vision-language benchmark covering Taiwan attractions, daily life, tables and diagrams, and university entrance exam questions, used to support the claim of vision understanding strength in Section 6.4.
    The paper states 'We plan to release TMMBench in the near future' (Section 6.4), so the only evidence of its validity is the paper's own description. No external audit or independent use is possible yet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Breeze 2 Herd of Models: Traditional Chinese LLMs Based on Llama with Vision-Aware and Function-Calling Capabilities." pith.science (2026). https://pith.science/paper/X4UIW47E

@misc{pith2026250113921,
  author       = {Pith},
  title        = {Pith review of: The Breeze 2 Herd of Models: Traditional Chinese LLMs Based on Llama with Vision-Aware and Function-Calling Capabilities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X4UIW47E}},
  note         = {Machine review of arXiv:2501.13921}
}
read the original abstract

Llama-Breeze2 (hereinafter referred to as Breeze2) is a suite of advanced multi-modal language models, available in 3B and 8B parameter configurations, specifically designed to enhance Traditional Chinese language representation. Building upon the Llama 3.2 model family, we continue the pre-training of Breeze2 on an extensive corpus to enhance the linguistic and cultural heritage of Traditional Chinese. In addition to language modeling capabilities, we significantly augment the models with function calling and vision understanding capabilities. At the time of this publication, as far as we are aware, absent reasoning-inducing prompts, Breeze2 are the strongest performing models in Traditional Chinese function calling and image understanding in its size class. The effectiveness of Breeze2 is benchmarked across various tasks, including Taiwan general knowledge, instruction-following, long context, function calling, and vision understanding. We are publicly releasing all Breeze2 models under the Llama 3.2 Community License. We also showcase the capabilities of the model running on mobile platform with a mobile application which we also open source.

Figures

Figures reproduced from arXiv: 2501.13921 by the authors.

Figure 1
Figure 1. Breeze2 Architecture. Breeze2 utilizes the widely-adopted “ViT-MLP-LLM” paradigm, which integrates a pre-trained InternViT-300M-448px with LLMs of various sizes through a vision￾to-language MLP projector. Specifically, it combines the InternViT-300M-448px with Llama 3.2 3B and Llama 3.1 8B models. This architecture leverages the strengths of ViT, MLP, and LLM to process both visual and textual inputs effectively. 3 … view at source ↗
Figure 2
Figure 2. Training Process of Breeze2 models. Breeze2 models were trained with a multi-stage training approach, starting from the Llama 3.2 3B and Llama 3.1 8B base models. The process starts with extended text-to-text pretraining to enhance Traditional Chinese language understanding. It is followed by vision-alignment pretraining to integrate visual features, and post-training for refining text and visual instruction tuning … view at source ↗
Figure 3
Figure 3. Image for demo showing a prize distribution table. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Overview of the App Interface Design. This figure illustrates the main user interaction [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 10 canonical work pages

  1. [1]

    Coyo-700m: Image-text pair dataset

    Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022

    2022

  2. [2]

    Taitravel, 2024

    Yu-Ying Chang. Taitravel, 2024. URL https://huggingface.co/datasets/benchang1110/TaiVision-TaiTravel

    work page 2024

  3. [3]

    Allava: Harnessing gpt4v-synthesized data for a lite vision-language model, 2024 a

    Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model, 2024 a

    work page 2024

  4. [4]

    Sharegpt4v: Improving large multi-modal models with better captions, 2023

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. URL https://arxiv.org/abs/2311.12793

    arXiv 2023

  5. [5]

    Enhancing Function-Calling Capabilities in LLMs: Strategies for Prompt Formats, Data Integration, and Multilingual Translation

    Yi-Chang Chen, Po-Chun Hsu, Chan-Jan Hsu, and Da-shan Shiu. Enhancing function-calling capabilities in llms: Strategies for prompt formats, data integration, and multilingual translation. arXiv preprint arXiv:2412.01130, 2024 b . URL https://arxiv.org/abs/2412.01130

    work page Pith review arXiv 2024

  6. [6]

    How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites, 2024 c

    Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu...

    arXiv 2024

  7. [7]

    ICDAR2019 Robust Reading Challenge on Arbitrary-Shaped Text (RRC-ArT)

    Chee-Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaitao Zhang, Junyu Han, Errui Ding, Jingtuo Liu, Dimosthenis Karatzas, Chee Seng Chan, and Lianwen Jin. Icdar2019 robust reading challenge on arbitrary-shaped text (rrc-art), 2019. URL https://arxiv.org/abs/1909.07145

    work page Pith review arXiv 2019

  8. [8]

    Unsupervised cross-lingual representation learning at scale

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm \'a n, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. Unsupervised cross-lingual representation learning at scale. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault, editors, Proceedings of the 58th Annual Meeting of the ...

    doi:10.18653/v1/2020 2020

Show all 37 references
  1. [9]

    Documenting large webtext corpora: A case study on the colossal clean crawled corpus

    Jesse Dodge, Maarten Sap, Ana Marasovi \'c , William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. Documenting large webtext corpora: A case study on the colossal clean crawled corpus. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and S...

    2021

  2. [10]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  3. [11]

    Making the V in VQA matter: Elevating the role of image understanding in V isual Q uestion A nswering

    Yash Goyal, Tejas Khot, Douglas Summers - Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in V isual Q uestion A nswering. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  4. [12]

    Advancing the evaluation of traditional chinese language models: Towards a comprehensive benchmark suite, 2023

    Chan-Jan Hsu, Chang-Le Liu, Feng-Ting Liao, Po-Chun Hsu, Yi-Chang Chen, and Da shan Shiu. Advancing the evaluation of traditional chinese language models: Towards a comprehensive benchmark suite, 2023

    2023

  5. [13]

    Breeze-7b technical report

    Chan-Jan Hsu, Chang-Le Liu, Feng-Ting Liao, Po-Chun Hsu, Yi-Chang Chen, and Da-Shan Shiu. Breeze-7b technical report. arXiv preprint arXiv:2403.02712, 2024. URL https://arxiv.org/abs/2403.02712

    2024 arXiv

  6. [14]

    Ocr-free document understanding transformer, 2022

    Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer, 2022. URL https://arxiv.org/abs/2111.15664

    2022 arXiv

  7. [15]

    The stack: 3 tb of permissively licensed source code

    Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Muñoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. The stack: 3 tb of permissively licensed source code. Preprint, 2022

    2022

  8. [16]

    Oleksii Kuchaiev, Jason Li, Huyen Nguyen, Oleksii Hrinchuk, Ryan Leary, Boris Ginsburg, Samuel Kriman, Stanislav Beliaev, Vitaly Lavrukhin, Jack Cook, Patrice Castonguay, Mariya Popova, Jocelyn Huang, and Jonathan M. Cohen. Ne M o: a toolkit for building AI applications using ...

    2019 arXiv

  9. [17]

    What matters when building vision-language models?, 2024

    Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models?, 2024

    2024

  10. [18]

    Function calling leaderboard for zhtw, 2024

    Liang-Chieh Lee, Cheng-Wei Lin, Pei-Chen Ho, Chien-Yu Yu, Yi-Chang Chen, and Da-Shan Shiu. Function calling leaderboard for zhtw, 2024. URL https://github.com/mtkresearch/function-calling-leaderboard-for-zhtw

    2024

  11. [19]

    Slimorca: An open dataset of gpt-4 augmented flan reasoning traces, with verification, 2023

    Wing Lian, Guan Wang, Bleys Goodson, Eugene Pentland, Austin Cook, Chanvichet Vong, and "Teknium". Slimorca: An open dataset of gpt-4 augmented flan reasoning traces, with verification, 2023. URL https://https://huggingface.co/Open-Orca/SlimOrca

    2023

  12. [20]

    Fineweb-zhtw: Scalable curation of traditional chinese text data from the web, 2024

    Cheng-Wei Lin, Wan-Hsuan Hsieh, Kai-Xin Guan, Chan-Jan Hsu, Chia-Chen Kuo, Chuan-Lin Lai, Chung-Wei Chung, Ming-Jen Wang, and Da-Shan Shiu. Fineweb-zhtw: Scalable curation of traditional chinese text data from the web, 2024. URL https://arxiv.org/abs/2411.16387

    2024 arXiv

  13. [21]

    Taiwan llm: Bridging the linguistic divide with a culturally aligned language model

    Yen-Ting Lin and Yun-Nung Chen. Taiwan llm: Bridging the linguistic divide with a culturally aligned language model. arXiv preprint arXiv:2311.17487, 2023

    2023 arXiv

  14. [22]

    Mmc: Advancing multimodal chart understanding with large-scale instruction tuning, 2024 a

    Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning, 2024 a . URL https://arxiv.org/abs/2311.10774

    2024 arXiv

  15. [23]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 34892--34916. Curran Associates, Inc., 2023....

    2023

  16. [24]

    Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets

    Zuxin Liu, Thai Hoang, Jianguo Zhang, Ming Zhu, Tian Lan, Shirley Kokane, Juntao Tan, Weiran Yao, Zhiwei Liu, Yihao Feng, et al. Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets. arXiv preprint arXiv:2406.18518, 2024 b

    2024 arXiv

  17. [25]

    Le, Barret Zoph, Jason Wei, and Adam Roberts

    Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V. Le, Barret Zoph, Jason Wei, and Adam Roberts. The flan collection: Designing data and methods for effective instruction tuning, 2023

    2023

  18. [26]

    Orca: Progressive learning from complex explanation traces of gpt-4, 2023

    Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Awadallah. Orca: Progressive learning from complex explanation traces of gpt-4, 2023

    2023

  19. [27]

    Kosmos-2: Grounding multimodal large language models to the world, 2023

    Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world, 2023. URL https://arxiv.org/abs/2306.14824

    2023 arXiv

  20. [28]

    Laion-5b: An open large-scale dataset for training next generation image-text models, 2022

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion...

    2022 arXiv

  21. [29]

    Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning

    Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Re...

    2021

  22. [30]

    An improved traditional chinese evaluation suite for foundation model

    Zhi-Rui Tam and Ya-Ting Pai. An improved traditional chinese evaluation suite for foundation model. arXiv, 2023

    2023

  23. [31]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024

    Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Ziteng Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024....

    2024 arXiv

  24. [32]

    Coco-text: Dataset and benchmark for text detection and recognition in natural images, 2016

    Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images, 2016. URL https://arxiv.org/abs/1601.07140

    2016 arXiv

  25. [33]

    The all-seeing project v2: Towards general relation comprehension of the open world

    Weiyun Wang, Yiming Ren, Haowen Luo, Tiantong Li, Chenxiang Yan, Zhe Chen, Wenhai Wang, Qingyun Li, Lewei Lu, Xizhou Zhu, et al. The all-seeing project v2: Towards general relation comprehension of the open world. arXiv preprint arXiv:2402.19474, 2024

    2024 arXiv

  26. [34]

    Patil, Ion Stoica, and Joseph E

    Fanjia Yan, Huanzhi Mao, Charlie Cheng-Jie Ji, Tianjun Zhang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. Berkeley function calling leaderboard. https://gorilla.cs.berkeley.edu/blogs/8_berkeley_function_calling_leaderboard.html, 2024

    2024

  27. [35]

    Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A m...

    2024

  28. [36]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36, 2024

    2024

  29. [37]

    A high-quality text-rich image instruction tuning dataset via hybrid instruction generation, 2024

    Shijie Zhou, Ruiyi Zhang, Yufan Zhou, and Changyou Chen. A high-quality text-rich image instruction tuning dataset via hybrid instruction generation, 2024. URL https://arxiv.org/abs/2412.16364

    2024 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.