AbacusResearch/Jallabi-34B
AbacusResearch/Jallabi-34B is a 34 billion parameter language model derived from the LLaVA-v1.6-34B architecture, specifically its Llama-only weights. This model, developed by AbacusResearch, removes the Clip encoder part, making it suitable for text-only applications and loadable via LlamaForCausalLM. It demonstrates strong performance across various reasoning and language understanding benchmarks, including MMLU and HellaSwag, making it a capable foundation for general-purpose language tasks.
Loading preview...
AbacusResearch/Jallabi-34B Overview
AbacusResearch/Jallabi-34B is a 34 billion parameter language model, specifically the Llama-only weights extracted from the liuhaotian/llava-v1.6-34b architecture. This model has had its Clip encoder removed, making it a text-focused model designed to be loaded using LlamaForCausalLM.
Key Characteristics & Performance
Derived from the LLaVA framework, Jallabi-34B inherits its foundational language capabilities. Its licensing is indirectly linked to NousResearch/Nous-Hermes-2-Yi-34B.
Performance on the Open LLM Leaderboard indicates strong general language understanding and reasoning:
- Average Score: 70.73
- MMLU (5-Shot): 76.40
- HellaSwag (10-Shot): 83.81
- AI2 Reasoning Challenge (25-Shot): 66.04
- Winogrande (5-shot): 81.45
- GSM8k (5-shot): 65.20
Additional evaluations on a separate leaderboard show:
- Average Score: 25.97
- BBH (3-Shot): 43.62
- MMLU-PRO (5-shot): 40.91
Use Cases
This model is suitable for developers seeking a robust 34B parameter Llama-based model for various text-generation and understanding tasks, particularly those requiring strong reasoning and general knowledge. Its optimized structure, without the vision component, makes it efficient for purely language-based applications.