pankajmathur/model_007_preview
pankajmathur/model_007_preview is a 69 billion parameter Llama2-based model developed by Pankaj Mathur, designed for both explanatory and instructional tasks. This hybrid model, with a 32768-token context length, is fine-tuned on a diverse set of datasets including Open-Platypus, Alpaca, WizardLM, and Orca, making it suitable for a wide range of general-purpose instruction-following applications. It achieves a total average score of 0.7249 on key benchmarks like ARC Challenge, HellaSwag, MMLU, and TruthfulQA.
Loading preview...
Model Overview
pankajmathur/model_007_preview is a 69 billion parameter Llama2-based model, serving as a preview version of psmathur/model_007. Developed by Pankaj Mathur, this model is characterized by its hybrid explain + instruct style, making it versatile for various prompt formats.
Key Capabilities
- Instruction Following: Fine-tuned on a comprehensive collection of datasets including Open-Platypus, Alpaca, WizardLM, Dolly-V2, Dolphin Samples, Orca_minis_v1, Alpaca_orca, and WizardLM_orca, enabling robust instruction adherence.
- Dual Prompt Formats: Supports both Orca and Alpaca prompt formats, providing flexibility for integration into different systems.
- Performance: Achieves competitive results on standard benchmarks, with an average score of 0.7249 across ARC Challenge (0.7142), HellaSwag (0.8731), MMLU (0.6858), and TruthfulQA (0.6265) according to EleutherAI's Language Model Evaluation Harness.
Good For
- General-purpose AI assistants: Its hybrid nature and broad training data make it suitable for a wide array of conversational and instructional tasks.
- Research and Development: Ideal for developers and researchers exploring Llama2-based models with diverse instruction-tuning approaches.
- Applications requiring both explanation and direct instruction: The model's design caters to scenarios where detailed explanations and precise instructions are needed.