AceInstruct-7B: Versatile Instruction-Tuned Model
AceInstruct-7B is a 7.6 billion parameter model from NVIDIA's AceInstruct family, fine-tuned on Qwen2.5-Base using general SFT datasets. Unlike the specialized AceMath-Instruct, AceInstruct is designed for broad applicability across coding, mathematics, and general knowledge tasks.
Key Capabilities & Performance
- Versatile Performance: Achieves strong results across coding (HumanEval, MBPP), mathematics (GSM8K, MATH), and general knowledge (MMLU, MMLU Pro) benchmarks.
- Comparable to Qwen2.5-Instruct: While AceInstruct-1.5B outperforms its Qwen2.5 counterpart, AceInstruct-7B demonstrates performance similar to Qwen2.5-7B-Instruct across various benchmarks.
- Coding: Scores 85.37 on HumanEval and 74.32 on MBPP.
- Mathematics: Achieves 93.10 on GSM8K and 76.40 on MATH.
- General Knowledge: Scores 74.68 on MMLU and 54.50 on MMLU Pro.
Training & Resources
AceInstruct models are fine-tuned using the same general SFT datasets as AceMath-Instruct, emphasizing a balanced approach to diverse tasks. Further details are available on the NVIDIA research website and in the associated paper.
License
This model is intended for non-commercial use only, subject to the Creative Commons Attribution: Non-Commercial 4.0 International license and OpenAI's Terms of Use for data generated by OpenAI.