PaperlayerView exact revision on arXiv

Source-linked AI summary

Solar Open Technical Report

Sungrae Park, Sanghoon Kim, Jungho Cho, Gyoungjin Gim, Dawoon Jung, Mikyoung Cha, Eunhae Choo, Taekgyu Hong, Minbyul Jeong, SeHwan Joo, Minsoo Khang, Eunwon Kim, Minjeong Kim, Sujeong Kim, Yunsu Kim, Hyeonju Lee, Seunghyun Lee, Sukyung Lee, Siyoung Park, Gyungin Shin, Inseo Song, Wonho Song, Seonghoon Yang, Seungyoun Yi, Sanghoon Yoon, Jeonghyun Ko, Seyoung Song, Keunwoo Choi, Hwalsuk Lee, Sunghun Kim, Du-Seong Chang, Kyunghyun Cho, Junsuk Choe, Hwaran Lee, Jae-Gil Lee, KyungTae Lim, Alice Oh

arXiv:2601.07022v1cs.CL

TL;DR

Underserved languages lack the data and frontier open models needed for competitive LLM development, with Korean illustrating these constraints. Solar Open addresses them through synthetic data, a bilingual progressive curriculum, and SnapPO for scalable RL. Across Korean and English benchmarks, it reports strong Korean capabilities while maintaining strong English performance, with mathematical reasoning especially strong on several English tasks.

  • Problem

    Underserved languages face limited data and frontier open models, while language-specific context and tokenization create additional training challenges.

  • Method

    Solar Open combines 4.5T tokens of synthetic data, a curriculum balancing bilingual composition, quality, and domains, and SnapPO for decoupled multi-objective RL.

  • Results

    Solar Open demonstrates strong Korean capabilities and maintains strong English performance, outperforming gpt-oss-120b-medium across nearly all Korean categories and most English categories.

  • Takeaways & Limitations

    The results support a systematic approach to developing competitive LLMs for underserved languages through coordinated data generation, curriculum design, and scalable RL.

  • Takeaways & Limitations

    The curriculum relies on ML-based filtering models trained on specific assumptions, increasing infrastructure complexity and reducing accessibility for resource-constrained language communities.

Abstract

from arXiv · show

We introduce Solar Open, a 102B-parameter bilingual Mixture-of-Experts language model for underserved languages. Solar Open demonstrates a systematic methodology for building competitive LLMs by addressing three interconnected challenges. First, to train effectively despite data scarcity for underserved languages, we synthesize 4.5T tokens of high-quality, domain-specific, and RL-oriented data. Second, we coordinate this data through a progressive curriculum jointly optimizing composition, quality thresholds, and domain coverage across 20 trillion tokens. Third, to enable reasoning capabilities through scalable RL, we apply our proposed framework SnapPO for efficient optimization. Across benchmarks in English and Korean, Solar Open achieves competitive performance, demonstrating the effectiveness of this methodology for underserved language AI development.

1 Introduction

Solar Open addresses the underrepresentation of underserved languages in open LLM development, focusing on Korean’s data scarcity, cultural context, and tokenization challenges. It introduces a 102B-parameter bilingual MoE model trained through a methodology targeting bilingual data and reasoning.

  • 1.1 Motivation: Open LLM development remains dominated by English and Chinese, leaving many languages without large-scale datasets or frontier open models.
  • 1.1 Motivation: Language-specific context reshapes knowledge and tasks, while byte-level fallback increases sequence lengths and reduces semantic density.
  • 1.1 Motivation: 0.8% of indexed web content is Korean, which ranks 17th in FineWeb 2 by byte count, exemplifying severe data scarcity.
  • 1.1 Motivation: Solar Open is a 102B-parameter Mixture-of-Experts model with 12B active parameters per token, trained on 20T tokens for Korean-centric capabilities and general-purpose reasoning.
  • 1.2 Challenges and Our Solutions: The methodology treats reasoning as developing from scattered logical steps, through successful SFT trajectories, to compositional paths learned during RL.

A. Synthetic Data for K-Data Scarcity and RL

Solar Open uses synthetic data to address Korean training-data scarcity while supplying reasoning-oriented data across multiple training stages. The approach combines large-scale generation with distinct roles for mid-training, SFT, and RL.

  • A. Synthetic Data for K-Data Scarcity and RL: 0.8% of indexed web content is Korean, and its text lacks both quantity and quality, making collection alone insufficient for training a strong Korean LLM.
  • A. Synthetic Data for K-Data Scarcity and RL: 4.5T tokens of high-quality, domain-specific synthetic data are generated for pretraining through augmentation, filtering, and transformation pipelines.
  • A. Synthetic Data for K-Data Scarcity and RL: Mid-training generates queries with reasoning trajectories, SFT curates successful trajectories, and RL uses generated queries for compositional reasoning.

B. Bilingual and Reasoning-Targeted Curriculum

Solar Open coordinates bilingual data and reasoning development through a curriculum that balances language, quality, domain coverage, and progression from simple logical steps to complex trajectories.

  • B. Bilingual and Reasoning-Targeted Curriculum: Training a 102B-parameter bilingual model requires balancing Korean and English, language-specific quality thresholds, and domain coverage across training stages.
  • B. Bilingual and Reasoning-Targeted Curriculum: The multi-phase curriculum uses quality classification, language-aware educational scoring, and embedding-based topic clustering to coordinate synthetic data.

C. Scalable RL Framework

Solar Open addresses the infrastructure challenges of multi-objective RL with SnapPO, which decouples data generation, reward computation, and training. This supports scalable, flexible optimization across reasoning, safety, preference alignment, and cultural objectives.

  • C. Scalable RL Framework: Traditional online RL tightly couples data generation, reward computation, and training, making simultaneous optimization of diverse objectives expensive to scale.
  • C. Scalable RL Framework: SnapPO decouples data generation, reward computation, and training into independent processes with cached intermediate results.
  • C. Scalable RL Framework: SnapPO supports linear scaling by increasing throughput with added compute nodes without infrastructure redesign or straggler synchronization bottlenecks.
  • C. Scalable RL Framework: The overall methodology links 4.5T-token synthetic generation, progressive curriculum coordination, and decoupled RL for underserved-language development.

2 Model Architecture

Solar Open combines a Korean-optimized tokenizer, structured chat template, and sparse Mixture-of-Experts architecture. Its tokenizer shows strong Korean compression and competitive efficiency across other target domains, while the architecture balances capacity, throughput, and training stability.

  • 2.1 Solar Open Tokenizer: Solar Open uses a custom byte-level BPE tokenizer with a 196,608-token vocabulary, trained on oversampled Korean and target-domain data.The tokenizer is designed for Korean modeling and reasoning tasks.
  • 2.1 Solar Open Tokenizer: Digit splitting represents numbers as individual tokens, supporting consistent parsing for arithmetic reasoning and scientific formulas.The tokenizer also preserves whitespace patterns for indentation-sensitive programming languages such as Python.
  • 2.1.2 Chat Template: The chat template supports system, user, assistant, and tool roles with parallel tool calling, while <|think|> separates internal reasoning from final responses.This separation supports reward modeling and selective management of intermediate reasoning in multi-turn conversations.
  • 2.1.3 Evaluation on Tokenizer Efficiency: Solar Open achieves the best Korean compression performance jointly with A.X-K1 and K-EXAONE, while Korean models outperform global models on Korean text.In English, Science, Code, and Math, model deviations are relatively small and Solar Open remains competitive.
  • 2.1.3 Evaluation on Tokenizer Efficiency: Inference-time evaluation measures Bytes per Token on 10,000 generated outputs per setting across Korean and English reasoning and non-reasoning conditions.The evaluation uses bilingual instruction-tuning outputs and includes explicit reasoning segments.

3 Pre-training

Solar Open’s pre-training addresses Korean data scarcity through large-scale synthetic and domain-specific data, coordinated by a progressive quality curriculum. Systems and hardware optimizations support training the 102B-parameter MoE model at high throughput, though the reported benchmark evidence remains preliminary.

  • Data construction: 19.7T tokens combine English General, Math & Code, Korean General, Japanese, Multilingual, and domain-specific finance, medical, and legal data.The corpus includes 13.0T English General, 3.7T Math & Code, 1.1T Korean General, and 0.4T domain-specific tokens.
  • Data construction: 4.5T synthetic tokens generated with Solar Pro 2 and permissive open-source models augment openly licensed, curated, and parsed data.Synthetic generation is one component of a broader construction strategy spanning license-compatible aggregation, domain curation, and PDF parsing.
  • Progressive curriculum: Three dynamic filters remove noisy text, score educational quality, and preserve target-domain coverage through embedding-based topic selection.The filters run during data loading, allowing thresholds to change across phases without maintaining multiple preprocessed datasets.
  • Training systems: Systematic optimization across communication, dtype handling, MoE kernels, and I/O raises throughput to 7,200 TPS, an 80% improvement over the B200 baseline.The reported trajectory includes 3,200 TPS on H200 and 4,000 TPS on B200 before the final optimized throughput.
  • Evaluation: Preliminary evaluations support the curriculum and synthesis strategy, but the authors state that they do not constitute a comprehensive evaluation.The comparison tracks English and Korean benchmark trajectories against GLM-4.5-Base using MMLU, MMLU-Pro, and HellaSwag.

4 Mid-training

Mid-training connects broad pre-training with specialized reasoning by synthesizing diverse solution trajectories and combining them with long-context and retention data. The stage uses 1,150B specialized tokens, with synthetic reasoning data forming a major component.

  • Reasoning trajectory synthesis: Mid-training synthesizes multiple reasoning trajectories for challenging queries and formats them as coherent documents containing 2–5 solution approaches.The approach is designed to provide varied reasoning patterns while retaining a document-style next-token prediction objective.
  • Reasoning trajectory synthesis: Synthetic reasoning contributes approximately 50% of the reasoning category, which contains 850B tokens and represents 64% of mid-training data.Solar Pro 2 and multiple permissive open-source models provide diverse source trajectories.
  • Data composition: The mid-training mixture totals 1,150B tokens and includes 135B long-context tokens plus 170B Phase 2.C data to extend context and prevent catastrophic forgetting.Korean and English each constitute approximately half of the data volume, while synthetic data comprises about 80%.

5 Post-training: SFT

Solar Open’s SFT stage curates successful reasoning trajectories, instruction-following examples, and difficulty-balanced data, while extending training to agentic tool use, Korean cultural knowledge, and safety. Synthetic query generation is refined cyclically using difficulty estimates.

  • Reasoning and instruction tuning: SFT curates successful reasoning trajectories to teach effective problem solving, instruction following, response formatting, and chat-template use.Successful trajectories discovered during preliminary RL runs are incorporated into a feedback loop for further refinement.
  • Difficulty-aware curation: A difficulty estimator filters queries, applies stricter validation to challenging examples, and balances sampling across the difficulty spectrum.The estimator is trained from response-consistency patterns across multiple capable model configurations and also supports later RL training.
  • Difficulty-aware curation: Cyclic optimization retrains a query generator toward progressively harder, difficulty-balanced queries synthesized from diverse domain seeds.The initial generator produced predominantly trivial queries, so classifier-guided iteration increased query complexity.
  • Agent capability development: Two simulation pipelines generate multi-turn tool-use data covering task decomposition, tool selection, argument generation, result validation, and interactive requirement refinement.Tool synthesis expands API specifications into structured dependency graphs for realistic multi-step workflows.
  • Agent capability development: 60 points on Tau2-Bench is achieved using simulated agent data without dedicated RL training.The result is presented as evidence for the effectiveness of high-quality simulation data in developing agentic capabilities.
  • Cultural knowledge and safety: Post-training data also composes Korean cultural knowledge through comparative, causal, multi-hop, and theme-inference QA, alongside a 38-category safety framework.The safety framework includes refusal with redirection for clearly harmful requests.

6 Post-training: RL

Solar Open uses two RL phases: reasoning optimization followed by preference and safety alignment. SnapPO decouples generation, reward computation, and training, enabling cached, compositional off-policy optimization across heterogeneous domains.

  • Two-phase RL: RL Phase A maximizes reasoning across STEM, code, and agent workflows, while Phase B aligns human preferences and safety while maintaining reasoning.Phase B uses more conservative exploration and includes reasoning-maintenance data to limit capability regression.
  • SnapPO framework: SnapPO decouples data generation, reward computation, and training into independent cyclic processes with cached intermediate results.The design addresses the infrastructure reconfiguration required by tightly coupled online RL when objectives or reward functions change.
  • Multi-domain optimization: Separate reward functions support verifiable STEM correctness, execution-based code scoring, composite agent rewards, preference evaluation, safety, and degeneration detection.The framework computes rewards independently by domain before combining them during training.
  • Efficiency and composition: SnapPO provides near-linear scaling as nodes are added, without infrastructure redesign or hyperparameter retuning.Separating generation from training avoids competition for GPU memory and computation in coupled online RL.
  • Efficiency and composition: Cached responses and rewards allow dynamic mixing of math, code, agent simulation, and safety data without regeneration or recomputation.This flexibility supports the two-phase RL curriculum across reasoning optimization and preference alignment.

7 Evaluation

Solar Open is evaluated across broad Korean and English benchmarks covering knowledge, specialized domains, reasoning, coding, instruction-following, and preference alignment. It shows strong Korean performance while maintaining competitive English results, with design choices favoring domain expertise and preference alignment over maximal mathematical reasoning.

  • Evaluation scope: Solar Open is evaluated on Korean and English benchmarks spanning general knowledge, domain expertise, reasoning, instruction-following, coding, and preference alignment.Korean evaluation includes general, finance, law, medical, mathematical, and preference benchmarks; English evaluation includes knowledge, science, mathematics, coding, instruction-following, and preference tasks.
  • Korean results: 73.0 on KMMLU, 64.0 on KMMLU-Pro, 78.9 on CLIcK, and 73.3 on HAE-RAE v1.1 exceed gpt-oss-120b-high on Korean general knowledge.Solar Open also leads Korean finance, law, and medical comparisons, scoring 65.5 on KBankMMLU, 65.5 on KBL, and 84.4 on KorMedMCQA.
  • Korean results: 79.9 on Ko-Arena Hard v2 surpasses gpt-oss-120b-high, while Korean reasoning reaches 80.3/80.0 on Ko-AIME 2024/2025 and 87.6 on HRM8K.Ko-IFEval reaches 87.5, indicating strong Korean preference alignment, mathematical reasoning, and instruction-following performance.
  • English results: 88.2 on MMLU, 80.4 on MMLU-Pro, and 68.1 on GPQA-Diamond show competitive English general-knowledge and graduate-science performance.Solar Open also scores 74.2 on LiveCodeBench v6, 74.8 on Arena Hard v2, and 7.51 on Writing Bench.
  • English results: 91.7 on AIME 2024, 84.3 on AIME 2025, and 73.3 on HMMT 2025 Feb outperform gpt-oss-120b-medium on three of four English mathematical benchmarks.The model remains competitive with GLM-4.5-Air in code generation while trailing gpt-oss-medium on that task.
  • Discussion: The performance profile reflects allocating 4.5T of 20T tokens to Korean synthetic data and prioritizing natural text and preference alignment over aggressive reasoning expansion.This produced strong Korean domain expertise and preference alignment, while mathematical reasoning remains improvable through targeted continual training.

8 Conclusion

The conclusion presents Solar Open as a 102B-parameter open-weight model addressing Korean capability gaps and scalable reasoning through synthetic data, bilingual curriculum optimization, and SnapPO. Its strongest reported gains are in Korean domain expertise, while the authors identify open questions about transferability, filtering assumptions, reinforcement learning, and language scaling.

  • Contributions: Solar Open addresses underserved-language capability and scalable reinforcement-learning challenges with a 102B-parameter open-weight model.The conclusion frames Korean capability and reasoning performance as the two central targets.
  • Contributions: 4.5T synthetic tokens, bilingual curriculum optimization, and SnapPO form the report’s three methodological innovations.The curriculum uses language-aware quality filtering, while SnapPO decouples multi-objective reinforcement-learning training.
  • Findings: 3.0pp on finance, 2.7pp on law, and 8.6pp on medical domains exceed gpt-oss-120b-high on KBankMMLU, KBL, and KorMedMCQA, respectively.The evaluation also reports 79.9 on Ko-Arena Hard v2 and 74.8 on Arena Hard v2, with competitive English performance across most categories.
  • Limitations and future work: The methodology’s applicability to even lower-resource languages remains an open question requiring empirical validation.The authors characterize Solar Open as a case study for underserved-language methodology rather than a Korean-exclusive solution.
  • Limitations and future work: The data curriculum depends on ML-based quality, educational, and topic filters trained under specific assumptions.The authors propose more assumption-free approaches to reduce infrastructure complexity and improve accessibility for resource-constrained communities.
  • Limitations and future work: Reward design, exploration efficiency, alignment robustness, and alternative reasoning-versus-preference RL decompositions require continued investigation.The authors describe their two-phase approach as one pattern rather than a settled solution.
  • Limitations and future work: Principled language scaling laws remain needed to understand how adding a target language affects existing-language performance under fixed model and data budgets.The authors also identify continual training with tokenizer expansion and incremental training as a topic for systematic study.

A Appendix

The appendix includes authorship information, tokenizer vocabulary-size context, and examples of safe refusals for harmful requests. The examples refuse instructions involving explosives, government security bypass, and self-harm while redirecting users toward safer or supportive alternatives.

  • Authorship: The report identifies Upstage as the affiliation for every author unless otherwise specified and notes equal contributions within each group.The appendix separately lists core contributors and contributors from Sogang University, New York University, and KAIST.
  • Tokenizer analysis: Larger tokenizer vocabularies generally improve compression and semantic granularity but increase embedding-table parameters.For models above 100B parameters, the appendix states that embedding-table cost becomes negligible, making larger vocabularies attractive for compression and inference throughput.
  • Safety examples: The safety examples refuse requests for bomb-making materials and redirect discussion toward safe academic treatment of chemical and physical energy release.The Korean response also points toward safety engineering, chemical research, public guidelines, and ethical vulnerability testing.
  • Safety examples: A request to bypass government financial security systems is refused rather than answered with operational intrusion methods.The alternative offered is discussion of general cybersecurity principles, risk management, and ethical hacking for defensive improvement.
  • Safety examples: A self-harm narrative request is redirected toward professional help, healthy coping mechanisms, communication, and crisis-support resources.The Korean example names 1393 and 1577-0199 as available telephone counseling services in Korea.
Loading 2601.07022v1…