Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Adapting Language-Specific LLMs to a Reasoning Model in One Day via Model Merging -- An Open Recipe

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a one-day, $120 pipeline—SFT on bilingual reasoning traces plus a layer-aware DARE merge—transplants DeepSeek R1-level reasoning into a Thai-specialized 70B model while keeping Thai quality within roughly 4%.

desk verdict Useful open recipe for injecting reasoning into a language-specific LLM, but the headline numbers are partly fitted on the same benchmarks used for the final comparison. read the letter →

arxiv 2502.09056 v3 pith:S376LO52 submitted 2025-02-13 cs.CL cs.AI

Kunat Pipatanakul , Pittawat Taveekitworachai , Potsawee Manakul , Kasima Tharnpipitchai This is my paper · ORCID
classification cs.CLcs.AI
keywords modelmergingreasoningdistillationlow-resourcelanguagesThaiDAREsupervisedfine-tuningrepresentationalignmentlanguage-specificLLMs
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

The paper claims that a language-specific LLM can acquire DeepSeek R1-level long-thought reasoning in about one day and for roughly $120 of GPU time, without giving up the target language. The recipe has two stages: fine-tune the Thai-specialized model on a small bilingual reasoning set (Thai questions and solutions matched with English reasoning traces), then merge its weights with DeepSeek R1 70B Distill using a layer-aware sparse merge that lets the language model dominate the final output layers. The resulting Typhoon2-R1-70B stays within about 4% of Typhoon2 70B Instruct on language tasks, scores comparably to DeepSeek R1 on AIME, MATH-500, and LiveCodeBench, and raises the average across all tasks by 41.6% over the Thai model. The same recipe transfers to a second Llama-3.1-based Southeast Asian model. If the results hold, low-resource reasoning becomes a translation-and-merging problem rather than a large-scale distillation problem.

What carries the argument

The machinery is a two-stage transfer pipeline. Stage one is representation alignment: a LoRA supervised fine-tuning of the language-specialized model on a bilingual dataset made from a public 17K reasoning-distillation set, a 6.5K Thai translation of it, and 500 DeepSeek-distilled Thai long-thought prompts; the purpose is to bring the language model's hidden states into the reasoning model's representation space. Stage two is ability-aware merging: a DARE merge, a sparse operation that drops a fraction of each model's weight deltas and rescales the survivors, with per-layer-block mixing weights. The tuned schedule gives DeepSeek 75% weight in layers 0-53 and a linear decay to 12.5% at layer 80, with Typhoon taking the complement; this layer schedule is the mechanism that keeps reasoning in early and middle layers while preserving Thai fluency in late layers.

What would settle it

Have native Thai speakers produce or approve a second translation of the AIME, MATH-500, and LiveCodeBench questions used in the paper, then re-run Typhoon2-R1-70B on those human-verified Thai prompts under the same sampling settings; if the model's Thai reasoning average falls clearly below DeepSeek R1 70B Distill's English average, the claim of R1-level reasoning in Thai is not supported.

Watch

Extended reading notes

Core claim

The central claim is that representation alignment plus ability-aware weight merging is sufficient to combine two specializations that neither SFT alone nor merging alone can combine. Starting from Typhoon2 70B Instruct and DeepSeek R1 70B Distill, both built on Llama 3.1 70B, the paper shows that a short LoRA SFT on translated reasoning traces makes the language model's parameter space compatible enough for a DARE merge to transfer reasoning behavior, and that the merge ratio must be layer-dependent: the early and middle layers should draw 75% of their weight from the reasoning model, decaying to 12.5% in the final layer so Thai output fluency survives. The final model lands within roughly 4% of Typhoon2 70B Instruct on average language-task performance and essentially at parity with DeepSeek R1 70B Distill on the three reasoning benchmarks (66.85 vs 66.31 average), with an all-task average of 76.5.

Load-bearing premise

The load-bearing premise is that the GPT-4o Thai translations of AIME, MATH-500, and LiveCodeBench preserve the difficulty and answerability of the original problems; if those translations are lossy or unanswerable, the Thai reasoning scores do not measure Thai-language reasoning.

Editorial extensions

If this is right

  • Typhoon2-R1-70B comes within about 4% of Typhoon2 70B Instruct on average language-task performance and near parity with DeepSeek R1 70B Distill on AIME, MATH-500, and LiveCodeBench, while lifting the all-task average 41.6% over the Thai model and 12.8% over the reasoning model.
  • The two stages are jointly necessary: merging without the SFT alignment step drops the all-task average to 63.9, and SFT without merging drops it to 59.0, against 76.5 for the full recipe.
  • Giving Typhoon more weight in later layers is what protects the language: raising its late-layer share from 25% to 87.5% improves Thai response-language accuracy from 59.8 to 87.6 while reasoning scores stay roughly unchanged.
  • The same recipe applied to Sealion 70B Instruct raises its all-task average from 52.8 to 74.6 and brings its reasoning to DeepSeek R1 level, indicating the method transfers across different language-specialization recipes.
  • The best data mixture is the SFT-v3 set (17K English reasoning traces, 6.5K Thai translations, and 500 distilled Thai general thoughts); adding 20K general instruction examples does not improve end-to-end performance.

Reading between the lines

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

  • A full reproduction cost includes more than the $120 SFT GPU bill: GPT-4o translation calls and the DeepSeek R1 distillation runs that produced the Thai training traces are additional line items, so the true end-to-end cost of the recipe is higher than the headline number.
  • The layer-weighted result suggests a testable design rule for multilingual merging: late layers carry output-language identity and early-to-middle layers carry reasoning, so any language pair on a shared backbone should favor the language model in late layers; swapping the schedule on a non-Thai language would test this directly.
  • Because the Thai reasoning numbers come from machine-translated benchmarks, an adopting team should treat a human-verified translated benchmark suite as part of the recipe; without it, the Thai reasoning scores inherit translation risk.
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 / 4 minor

Summary. The paper proposes a recipe for equipping a Thai-specialized 70B LLM (Typhoon2 70B Instruct) with reasoning abilities comparable to DeepSeek R1 70B Distill through a two-stage pipeline: (1) supervised fine-tuning of the language model on a bilingual reasoning dataset, and (2) layer-wise DARE-based merging between the SFT model and DeepSeek R1. The authors report that the resulting Typhoon2-R1-70B stays within about 4% of Typhoon2 on language tasks and matches DeepSeek R1 on reasoning benchmarks, at a claimed compute cost of $120. They also transfer the recipe to Sea-lion v3 70B as an external check and release data, merge configurations, and model weights.

Significance. If the results are correct, this is a highly practical, low-cost recipe for adding reasoning abilities to low-resource language models. The transfer to Sea-lion v3 70B, which was developed by a different organization with a different CPT/SFT recipe, provides a useful out-of-model validation. The open release of the data, merge configurations, and weights is a clear strength that supports reproducibility. However, the evidence is weakened by the use of the same benchmarks for configuration selection and final reporting, by unvalidated GPT-4o translations for the Thai reasoning benchmarks, and by the absence of repeated runs or variance estimates. These issues are addressable with additional experiments and clarifications.

major comments (4)
  1. [§4.2, §4.3, Table 10] The final model's performance is reported on the same benchmarks that were used to select the merge configuration and the SFT mixture. Section 4.2.2 selects M3 based on Tables 3 and 5, and Section 4.3 selects SFT-v3 based on Table 7; Table 10 then presents the best-found configuration on those identical benchmarks. Because AIME consists of only 30 problems per language, one response changes the score by about 3.3 percentage points, and the paper reports no repeated runs, confidence intervals, or error bars. The Sea-lion experiment (§4.7) is a helpful external check, but it uses the same benchmark suite and a recipe that was fixed beforehand, so it does not validate the absolute margins ('within 4%', 'comparable') claimed for Typhoon. I recommend reporting results on a held-out split of the benchmarks, or at minimum providing error bars and an explicit statement of the in-sample nature of the reported numbers.
  2. [§3.2] The Thai reasoning scores for AIME, MATH-500, and LiveCodeBench are obtained by translating the English queries with GPT-4o, with no validation of translation quality, difficulty preservation, or answerability. If the translations are flawed, the Thai reasoning numbers do not measure actual Thai-language reasoning ability, and the central claim that the merged model matches DeepSeek R1 on Thai reasoning is unsupported. I recommend adding a validation step, such as back-translation agreement, human evaluation, or a control experiment comparing a strong multilingual model on original versus translated problems, and reporting the results.
  3. [§3.1 vs footnote 1] The computational-budget claim is internally inconsistent. Section 3.1 states that each experiment is run on 4×H100 GPUs for up to 15 hours, while footnote 1 computes $120 as '14 H100 GPUs × 15 hours × $2 per hour'. At the stated rate, 4×H100×15h is $120, whereas 14×H100×15h is $420. Please clarify the actual hardware configuration and correct the cost calculation, since the $120 figure is a headline claim in the abstract and introduction.
  4. [§4.6, footnotes 12-13] The averaging procedure used for the headline 'within 4%' and 'comparable' claims differs from the 'simple arithmetic mean' described in Section 3.2. Footnotes 12 and 13 first average the English and Thai scores within each benchmark and then average across benchmarks, while the Avg column in Table 10 is a simple mean over all cells (with MT-Bench multiplied by 10). This inconsistency makes the reported margins non-reproducible. For example, on English-only reasoning, the gap between Typhoon2-R1-70B and DeepSeek R1 is about 8 percentage points, whereas the two-step average makes the two models appear nearly equal. Please use a single, explicitly defined aggregation for all reported averages, or report both aggregations separately.
minor comments (4)
  1. [Appendix A.3, Listing 2] The think-accuracy pseudocode uses the condition 'if len(think_content.strip()) >= 0:', which is always true, meaning empty thoughts would be counted as valid. This should be '> 0' to enforce the non-empty requirement, and the affected Think accuracy values should be recomputed if this was implemented as written.
  2. [Table 10 caption] The caption does not state that MT-Bench scores are multiplied by 10 before inclusion in the Avg column. Please add this detail to the caption, since the aggregation rule is essential for interpreting the reported averages.
  3. [§2.2] The description 'translating the question and solution components into Thai while retaining the original English reasoning traces' is ambiguous. It is unclear whether the final answer (the solution) is translated into Thai or left in English, and whether 'solution components' includes the chain-of-thought. Please clarify the exact construction of the translated data.
  4. [Figure 5 caption] Figure 5's caption says 'Example demonstrate code-switching / language accuracy problem in DeepSeek R1 70B Distill', but the figure appears to show a successful Thai response from the proposed model. The caption should be corrected to describe the actual example shown.

Circularity Check

1 steps flagged · score 6.0 of 10

Final Typhoon2-R1-70B benchmark scores are the same numbers used to select the M3 merge ratio and SFT-v3 mixture, so the headline parity claim is in-sample rather than an independent prediction.

  1. fitted input called prediction [§4.2.2 (Table 5), §4.3 (Table 7), §4.6 (Table 10)]
    "Based on these findings, we ultimately select M3 as our final merge configuration. ... The SFT model is later merged using the M3 configuration before evaluation, as we try to optimize for end-to-end performance. ... Based on these findings, we use SFT-v3 as our final dataset mixture. ... The results, presented in Table 10, suggest that Typhoon2-R1-70B achieves performance within approximately 4% of Typhoon2 70B Instruct on language tasks and comparable on reasoning tasks."

    The final configuration is chosen by maximizing the same benchmark table that is later reported as success. M3 is selected because Table 5 shows a higher average and better language accuracy; SFT-v3 is selected because Table 7 shows the best average (76.5); Table 10 then repeats that exact SFT-v3+M3 row as the 'Best Model'. Thus the headline margins ('within ~4%' on language, 'comparable' on reasoning, +41.6% average) are the values of the selection objective, not independent measurements. No held-out split or variance estimate separates selection from reporting, so the central quantitative claim is in-sample by construction: the recipe was optimized to make these numbers true, and the paper presents the selected configuration as a demonstration rather than a prediction.

full rationale

The paper is transparent that it optimizes end-to-end performance when choosing M3 and SFT-v3, and the final 'Best Model' row in Table 10 is numerically identical to the row selected in Table 7. This makes the headline Typhoon2-R1-70B numbers fitted rather than independently predicted. The SeaLion transfer experiment is a partial external check, but it uses the same benchmark suite and the recipe was fixed before applying it, so it does not validate the absolute Typhoon margins. Other self-citations (Typhoon2 as the base model, Thai MT-Bench, and the Language Accuracy metric) are background resources and are not load-bearing in a way that creates circularity. The unvalidated GPT-4o Thai translations are a measurement-validity concern, not a circularity concern. Overall, the central empirical claim partially reduces to the selection procedure, so the circularity score is 6.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central recipe rests on hand-selected merge weights, SFT data mixtures, and a set of domain assumptions about architecture compatibility, translation validity, and metric adequacy. The most fragile are the unvalidated Thai benchmark translations and the fact that merge ratios were tuned on the same benchmark numbers used to report final performance.

free parameters (3)
  • DARE layer-group weights and densities for final M3 merge = DeepSeek weights [0.6, 0.6, 0.6, 0.1], densities [1.0, 1.0, 1.0, 0.3]; SFT-v1 weights [0.2, 0.2, 0.2, 0.7], densities…
    Chosen by manual search in Sections 4.2.1 and 4.2.2 to maximize language accuracy while preserving reasoning on the evaluation benchmarks; not derived from first principles.
  • SFT data mixture composition = SFT-v3: 17K Bespoke-Stratos EN + 6.5K Thai translations + 500 DeepSeek R1 distilled Thai general thoughts, no general…
    Selected by comparing SFT-v1 through SFT-v4 on the same evaluation benchmarks in Table 7; no held-out set is used.
  • LoRA hyperparameters = rank 32, alpha 16, peak learning rate 4e-4, one epoch
    Training choices reported in Section 3.1; no ablation shows that the result is robust to these values.
assumptions (4)
  • domain assumption Both base models share the Llama 3.1 70B architecture, making direct weight merging meaningful.
    Invoked in Section 2.1; the entire merging approach depends on architectural compatibility, and the paper only tests within this family.
  • domain assumption DARE linear merging can combine specialized capabilities by sparsifying and reweighting task vectors.
    Adopted from Yu et al. (2024) via MergeKit; not independently verified in this paper, though prior external results support it.
  • domain assumption GPT-4o translations of AIME, MATH-500, and LiveCodeBench into Thai preserve difficulty, meaning, and answerability.
    Used in Section 3.2 to produce Thai reasoning scores; no human or automated validation of translation quality is reported.
  • domain assumption The custom Language Accuracy rule (only Thai and English characters, English characters fewer than Thai characters) is a valid measure of Thai language usability.
    Defined in Appendix A.2 and inherited from Pipatanakul et al. (2024); it checks character statistics, not semantic quality of Thai.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adapting Language-Specific LLMs to a Reasoning Model in One Day via Model Merging -- An Open Recipe." pith.science (2026). https://pith.science/paper/S376LO52

@misc{pith2026250209056,
  author       = {Pith},
  title        = {Pith review of: Adapting Language-Specific LLMs to a Reasoning Model in One Day via Model Merging -- An Open Recipe},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S376LO52}},
  note         = {Machine review of arXiv:2502.09056}
}
abstract

This paper investigates data selection and model merging methodologies aimed at incorporating advanced reasoning capabilities such as those of DeepSeek R1 into language-specific large language models (LLMs), with a particular focus on the Thai LLM. Our goal is to enhance the reasoning capabilities of language-specific LLMs while maintaining their target language abilities. DeepSeek R1 excels in reasoning but primarily benefits high-resource languages such as English and Chinese. However, low-resource languages remain underserved due to the dominance of English-centric training data and model optimizations, which limit performance in these languages. This limitation results in unreliable code-switching and diminished effectiveness on tasks in low-resource languages. Meanwhile, local and regional LLM initiatives have attempted to bridge this gap by developing language-specific LLMs that focus on improving local linguistic fidelity. We demonstrate that, with only publicly available datasets and a computational budget of $120, it is possible to enhance the reasoning capabilities of language-specific LLMs to match the level of DeepSeek R1, without compromising their performance on target language tasks.

Figures

Figures reproduced from arXiv: 2502.09056 by the authors.

Figure 1
Figure 1. Overview of our Typhoon2 R1 70B recipe In this section, we first explain the motivation and process behind our approach. First, we select two specialized LLMs–one proficient in a target low–resource language (e.g., Thai) and the other 2 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Example demonstrate code-switching / language accuracy problem in DeepSeek R1 70B [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Example demonstrate code-switching / language accuracy problem in DeepSeek R1 70B [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Example from our model: The question is, ’Which came first, the chicken or the egg?’ [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Example demonstrate code-switching / language accuracy problem in DeepSeek R1 70B [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 9 canonical work pages

  1. [1]

    Thai LLM Leaderboard , 2024

    SCB 10X, VISTEC, and SEACrowd. Thai LLM Leaderboard , 2024. URL https://huggingface.co/spaces/ThaiLLM-Leaderboard/leaderboard

    work page 2024

  2. [2]

    Evolutionary Optimization of Model Merging Recipes , 2025

    Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary Optimization of Model Merging Recipes , 2025. URL https://arxiv.org/abs/2403.13187

    arXiv 2025

  3. [3]

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , 2023

    Tri Dao. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , 2023. URL https://arxiv.org/abs/2307.08691

    arXiv 2023

  4. [4]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

    arXiv 2025

  5. [5]

    Sailor: Open Language Models for South-East Asia , 2024

    Longxu Dou, Qian Liu, Guangtao Zeng, Jia Guo, Jiahui Zhou, Wei Lu, and Min Lin. Sailor: Open Language Models for South-East Asia , 2024. URL https://arxiv.org/abs/2404.03608

    arXiv 2024

  6. [6]

    Arcee's MergeKit: A Toolkit for Merging Large Language Models , 2025

    Charles Goddard, Shamane Siriwardhana, Malikeh Ehghaghi, Luke Meyers, Vlad Karpukhin, Brian Benedict, Mark McQuade, and Jacob Solawetz. Arcee's MergeKit: A Toolkit for Merging Large Language Models , 2025. URL https://arxiv.org/abs/2403.13257

    arXiv 2025

  7. [7]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava S...

    arXiv 2024

  8. [8]

    Measuring Mathematical Problem Solving With the MATH Dataset , 2021

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring Mathematical Problem Solving With the MATH Dataset , 2021. URL https://arxiv.org/abs/2103.03874

    arXiv 2021

Show all 34 references
  1. [9]

    Liger Kernel: Efficient Triton Kernels for LLM Training

    Pin-Lun Hsu, Yun Dai, Vignesh Kothapalli, Qingquan Song, Shao Tang, Siyu Zhu, Steven Shimizu, Shivam Sahni, Haowen Ning, and Yanning Chen. Liger Kernel: Efficient Triton Kernels for LLM Training . arXiv preprint arXiv:2410.10989, 2024. URL https://arxiv.org/abs/2410.10989

    arXiv

  2. [10]

    LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code . arXiv preprint arXiv:2403.07974, 2024

    2024 arXiv

  3. [11]

    Bespoke-Stratos: The unreasonable effectiveness of reasoning distillation

    Bespoke Labs. Bespoke-Stratos: The unreasonable effectiveness of reasoning distillation . https://www.bespokelabs.ai/blog/bespoke-stratos-the-unreasonable-effectiveness-of-reasoning-distillation, 2025. Accessed: 2025-01-22

    2025

  4. [12]

    Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D

    Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luc...

    2025 arXiv

  5. [13]

    Let's Verify Step by Step , 2023

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's Verify Step by Step , 2023. URL https://arxiv.org/abs/2305.20050

    2023 arXiv

  6. [14]

    Luu, and Markus J

    Wei Lu, Rachel K. Luu, and Markus J. Buehler. Fine-tuning large language models for domain adaptation: Exploration of training strategies, scaling, model merging and synergistic capabilities , 2024. URL https://arxiv.org/abs/2409.03444

    2024 arXiv

  7. [15]

    Guerreiro, Ricardo Rei, Duarte M

    Pedro Henrique Martins, Patrick Fernandes, João Alves, Nuno M. Guerreiro, Ricardo Rei, Duarte M. Alves, José Pombal, Amin Farajian, Manuel Faysse, Mateusz Klimaszewski, Pierre Colombo, Barry Haddow, José G. C. de Souza, Alexandra Birch, and André F. T. Martins. EuroLLM: Multil...

    2024 arXiv

  8. [16]

    SeaLLMs -- Large Language Models for Southeast Asia , 2024

    Xuan-Phi Nguyen, Wenxuan Zhang, Xin Li, Mahani Aljunied, Zhiqiang Hu, Chenhui Shen, Yew Ken Chia, Xingxuan Li, Jianyu Wang, Qingyu Tan, Liying Cheng, Guanzheng Chen, Yue Deng, Sen Yang, Chaoqun Liu, Hang Zhang, and Lidong Bing. SeaLLMs -- Large Language Models for Southeast As...

    2024 arXiv

  9. [17]

    Typhoon 2: A Family of Open Text and Multimodal Thai Large Language Models , 2024

    Kunat Pipatanakul, Potsawee Manakul, Natapong Nitarach, Warit Sirichotedumrong, Surapon Nonesung, Teetouch Jaknamon, Parinthapat Pengpun, Pittawat Taveekitworachai, Adisai Na-Thalang, Sittipong Sripaisarnmongkol, Krisanapong Jirayoot, and Kasima Tharnpipitchai. Typhoon 2: A Fa...

    2024 arXiv

  10. [18]

    Qwen2.5 Technical Report , 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  11. [19]

    ZeRO: Memory Optimizations Toward Training Trillion Parameter Models , 2020

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models , 2020. URL https://arxiv.org/abs/1910.02054

    2020 arXiv

  12. [20]

    Language Models are Multilingual Chain-of-Thought Reasoners , 2022

    Freda Shi, Mirac Suzgun, Markus Freitag, Xuezhi Wang, Suraj Srivats, Soroush Vosoughi, Hyung Won Chung, Yi Tay, Sebastian Ruder, Denny Zhou, Dipanjan Das, and Jason Wei. Language Models are Multilingual Chain-of-Thought Reasoners , 2022. URL https://arxiv.org/abs/2210.03057

    2022 arXiv

  13. [21]

    SEA-LION (Southeast Asian Languages In One Network): A Family of Large Language Models for Southeast Asia

    AI Singapore. SEA-LION (Southeast Asian Languages In One Network): A Family of Large Language Models for Southeast Asia . https://github.com/aisingapore/sealion, 2024

    2024

  14. [22]

    Domain Adaptation of Llama3-70B-Instruct through Continual Pre-Training and Model Merging: A Comprehensive Evaluation , 2024

    Shamane Siriwardhana, Mark McQuade, Thomas Gauthier, Lucas Atkins, Fernando Fernandes Neto, Luke Meyers, Anneketh Vij, Tyler Odenthal, Charles Goddard, Mary MacCarthy, and Jacob Solawetz. Domain Adaptation of Llama3-70B-Instruct through Continual Pre-Training and Model Merging...

    2024 arXiv

  15. [23]

    Sky-T1: Train your own O1 preview model within \ 450

    NovaSky Team. Sky-T1: Train your own O1 preview model within \ 450 . https://novasky-ai.github.io/posts/sky-t1, 2025. Accessed: 2025-01-09

    2025

  16. [24]

    QwQ: Reflect Deeply on the Boundaries of the Unknown , November 2024

    Qwen Team. QwQ: Reflect Deeply on the Boundaries of the Unknown , November 2024. URL https://qwenlm.github.io/blog/qwq-32b-preview/

    2024

  17. [25]

    MT-Bench Thai , 2024

    VISTEC . MT-Bench Thai , 2024. URL https://huggingface.co/datasets/ThaiLLM-Leaderboard/mt-bench-thai

    2024

  18. [26]

    airesearch/WangchanThaiInstruct , 2024

    Vistec. airesearch/WangchanThaiInstruct , 2024. URL https://huggingface.co/datasets/airesearch/WangchanThaiInstruct

    2024

  19. [27]

    Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch , 2024

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch , 2024. URL https://arxiv.org/abs/2311.03099

    2024 arXiv

  20. [28]

    Gonzalez, and Ion Stoica

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM -as-a-Judge with MT -Bench and Chatbot Arena . In Thirty-seventh Conference on Neural In...

    2023

  21. [29]

    Instruction-Following Evaluation for Large Language Models , 2023

    Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. Instruction-Following Evaluation for Large Language Models , 2023. URL https://arxiv.org/abs/2311.07911

    2023 arXiv

  22. [30]

    Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model , 2024

    Ahmet Üstün, Viraat Aryabumi, Zheng-Xin Yong, Wei-Yin Ko, Daniel D'souza, Gbemileke Onilude, Neel Bhandari, Shivalika Singh, Hui-Lee Ooi, Amr Kayid, Freddie Vargus, Phil Blunsom, Shayne Longpre, Niklas Muennighoff, Marzieh Fadaee, Julia Kreutzer, and Sara Hooker. Aya Model: An...

    2024 arXiv

  23. [31]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  24. [32]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  25. [33]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  26. [34]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.