Einstein-v4-7B Overview
Einstein-v4-7B is a 7 billion parameter language model developed by Weyaxi, built upon the mistralai/Mistral-7B-v0.1 base model. It has undergone a full fine-tuning process for 1.5 epochs, utilizing a diverse set of datasets including alpaca, sharegpt, synthia-v1.3, cot_alpaca_gpt4, slimorca, and airoboros_3.2. The training was sponsored by sablo.ai and conducted using the axolotl framework.
Key Capabilities & Performance
This model is designed for general instruction following and conversational tasks. Its performance on the Open LLM Leaderboard (v1) shows an average score of 66.62, with notable results in:
- HellaSwag (10-Shot): 83.75
- Winogrande (5-shot): 76.24
- AI2 Reasoning Challenge (25-Shot): 64.68
On the Open LLM Leaderboard v2, it achieved an average of 16.73, with an IFEval (0-Shot) score of 47.08. The model uses a ChatML prompt template, supporting system, user, and assistant roles, and can be easily integrated using tokenizer.apply_chat_template().
Use Cases
Einstein-v4-7B is well-suited for applications requiring a capable 7B model for:
- General-purpose chat and conversational agents.
- Instruction-following tasks.
- Reasoning and common sense understanding, as indicated by its benchmark scores.
Quantized versions (GGUF, AWQ, Exl2) are also available for optimized deployment.