choco9966/Llama-2-7b-instruct-tuning is an instruction-tuned large language model based on the SOLAR-10.7B architecture, developed by Upstage. This 7 billion parameter model is fine-tuned for single-turn conversations, leveraging depth up-scaling (DUS) and integrating Mistral 7B weights. It demonstrates strong performance, particularly for fine-tuning applications, and is optimized for robust and adaptable instruction-following in single-turn interactions.
Loading preview...
Model Overview
This model, choco9966/Llama-2-7b-instruct-tuning, is an instruction-tuned version of Upstage's SOLAR-10.7B, a 10.7 billion parameter large language model. It utilizes a novel depth up-scaling (DUS) methodology, which involves integrating Mistral 7B weights into upscaled layers and continued pre-training. Despite its compact size, SOLAR-10.7B has shown strong performance, even surpassing some models with up to 30 billion parameters, including Mixtral 8x7B in certain evaluations.
Key Capabilities & Training
- Architecture: Based on SOLAR-10.7B, which employs depth up-scaling for enhanced performance.
- Instruction Fine-Tuning: Trained using state-of-the-art methods including Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO).
- Diverse Datasets: Fine-tuned on a mixture of datasets such as
c-s-ale/alpaca-gpt4-data,Open-Orca/OpenOrca, in-house generated Metamath data,Intel/orca_dpo_pairs, andallenai/ultrafeedback_binarized_cleaned. - Contamination-Free: Rigorous data contamination tests were conducted, showing results well below 0.9%, indicating the model is free from benchmark contamination.
Good For
- Single-Turn Conversations: The model is primarily fine-tuned for single-turn conversational tasks.
- Fine-Tuning: SOLAR-10.7B is highlighted as an ideal choice for further fine-tuning due to its robustness and adaptability.
- Research & Development: Suitable for researchers and developers exploring efficient yet powerful LLMs, especially those interested in the DUS methodology.