Overview

polymer/model-007-2-13b is a 13 billion parameter language model based on the Llama2 architecture, specifically a modified fork of psmathur/model_007_13b_v2. This version is prepared for training with the Hugging Face Transformers library, offering a readily usable base for further fine-tuning.

Key Capabilities

• Hybrid Explain + Instruct Style: The model is fine-tuned to handle both explanatory and instructional prompts effectively, making it versatile for various conversational and task-oriented applications.
• Diverse Training Data: It was trained on a comprehensive set of instruction-following datasets, including Open-Platypus, Alpaca, WizardLM, Dolly-V2, Dolphin Samples, and Orca variants, enhancing its ability to follow complex instructions.
• Performance Benchmarks: Evaluated using the EleutherAI Language Model Evaluation Harness, the model achieved notable scores:
  • ARC Challenge (acc_norm): 0.6314
  • HellaSwag (acc_norm): 0.8242
  • MMLU (acc_norm): 0.5637
  • TruthfulQA (mc2): 0.5127
  • Total Average: 0.6329

Usage and Limitations

This model supports common prompt formats like Orca and Alpaca, with example code provided for easy integration. Users should be aware that, like all large language models, it may occasionally produce inaccurate, biased, or offensive content. Its license is bound by the original Llama-2 model's restrictions.