SJ-Donald/SOLAR-10.7B-slerp

TEXT GENERATIONConcurrent Unit Cost:1Model Size:15BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jan 12, 2024License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SJ-Donald/SOLAR-10.7B-slerp is a 10.7 billion parameter language model created by SJ-Donald through a slerp merge of LDCC/LDCC-SOLAR-10.7B and upstage/SOLAR-10.7B-Instruct-v1.0. This model demonstrates strong performance across various benchmarks, including an average of 72.58 on the Open LLM Leaderboard and 56.93 on the Open-Ko-LLM-Leaderboard, indicating its capabilities in general reasoning and Korean language tasks. It is suitable for applications requiring a balanced performance across diverse linguistic and reasoning challenges.

Loading preview...

Model Overview

SJ-Donald/SOLAR-10.7B-slerp is a 10.7 billion parameter language model developed by SJ-Donald. It was created using the slerp (spherical linear interpolation) merge method via mergekit, combining two base models: LDCC/LDCC-SOLAR-10.7B and upstage/SOLAR-10.7B-Instruct-v1.0.

Key Capabilities & Performance

This model exhibits robust performance across a range of benchmarks:

  • Open LLM Leaderboard: Achieves an average score of 72.58, with notable results in:
    • AI2 Reasoning Challenge (25-Shot): 68.17
    • HellaSwag (10-Shot): 86.91
    • MMLU (5-Shot): 66.73
    • TruthfulQA (0-shot): 67.42
    • Winogrande (5-shot): 84.06
    • GSM8k (5-shot): 62.17
  • Open-Ko-LLM-Leaderboard: Demonstrates an average score of 56.93, indicating proficiency in Korean language tasks, including:
    • Ko-ARC: 53.58
    • Ko-HellaSwag: 62.03
    • Ko-MMLU: 53.31
    • Ko-TruthfulQA: 57.16
    • Ko-CommonGen V2: 58.56

When to Use This Model

SOLAR-10.7B-slerp is well-suited for use cases requiring a versatile language model with strong general reasoning and language understanding capabilities. Its performance on both general and Korean-specific benchmarks suggests it can be effectively applied in multilingual contexts or applications where a balanced performance across diverse tasks is critical. Developers can integrate it using the provided Hugging Face transformers library example.