megastudyedu/M-SOLAR-10.7B-v1.1-beta
M-SOLAR-10.7B-v1.1-beta is a large language model developed through a joint research effort by MegastudyEdu, Prediction, and Myice. This model is instruction fine-tuned using a diverse set of Korean datasets, including KOR-OpenOrca-Platypus-v2, KorQuAD 2.1, AIHUB technical science summary data, and in-house educational domain data. It is specifically designed to excel in tasks requiring text generation, summarization, and translation, with a particular focus on Korean language understanding and generation, and enhanced English-to-Korean translation capabilities.
Loading preview...
M-SOLAR-10.7B-v1.1-beta: Korean-centric Instruction Fine-tuned LLM
M-SOLAR-10.7B-v1.1-beta is a large language model developed collaboratively by MegastudyEdu, Prediction, and Myice. This model has undergone extensive instruction fine-tuning using a variety of Korean datasets, aiming to enhance its performance across several key areas.
Key Capabilities & Training Strategy
The model's training incorporates a diverse set of data and innovative strategies:
- Instruction Fine-Tuning: Utilizes datasets like kyujinpy/KOR-OpenOrca-Platypus-v2 and a subset of KorQuAD 2.1 for general instruction following.
- Scientific Text Generation: Leverages AIHUB technical science summary data to generate full scientific texts from summaries and clues, demonstrating advanced inference capabilities.
- Domain-Specific Knowledge: Incorporates in-house educational domain data from MegastudyEdu and Prediction, tailoring the model for educational applications.
- Self-Supervised Learning: Employs a unique "Random Split Generation" method where the model reconstructs original sentence order from shuffled text, enhancing its understanding of textual coherence.
- Enhanced Translation: Integrates AIHUB industrial information patent English-Korean data, specifically designed to improve English-to-Korean translation accuracy and capacity by combining multiple patent documents within a 4096-token limit.
- Additional Datasets: Further fine-tuned with HumanF-MarkrAI/WIKI_QA_Near_dedup and kyujinpy/KoCoT_2000 to broaden its knowledge and reasoning abilities.
Good For
- Korean Language Processing: Excels in tasks requiring deep understanding and generation of Korean text.
- Text Summarization and Expansion: Capable of generating detailed content from concise summaries, particularly in scientific and technical domains.
- English-to-Korean Translation: Offers improved translation quality for technical and patent-related content.
- Educational Applications: Benefits from specialized in-house data, making it suitable for educational content generation and analysis.
- Contextual Reasoning: The self-supervised training on sentence reordering helps in tasks requiring logical flow and coherence.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.