Overview
unaidedelf87777/wizard-mistral-v0.1 is a 7 billion parameter language model built upon the Mistral architecture. It has been fine-tuned by unaidedelf87777 using a meticulously cleaned version of the WizardLM Evol Instruct v2 196k dataset. A key characteristic of this model is the deliberate removal of most RLHF (Reinforcement Learning from Human Feedback) instances from its training data, which contributes to a less censored output compared to many other instruction-tuned models.
Key Capabilities
- Instruction Following: Designed to respond to a wide range of instructions based on its training on the Evol Instruct dataset.
- Less Censored Responses: Due to the reduced RLHF, the model tends to provide more direct and less restricted answers.
- Competitive Performance: Despite its focused training, it demonstrates competitive benchmark scores when evaluated against other prominent Mistral-7B fine-tunes.
Benchmarks
The model's performance was evaluated against several popular Mistral-7B fine-tunes across standard benchmarks. While not leading in all categories, unaidedelf87777/wizard-mistral-v0.1 achieves an average score of 64.18, with specific scores including 61.77 on ARC, 83.51 on HellaSwag, 63.99 on MMLU, and 47.46 on TruthfulQA. These results position it as a viable option within its class.
Use Cases
This model is particularly well-suited for applications requiring a general-purpose conversational AI that can handle diverse prompts with a tendency towards less constrained or filtered responses. It can be used for various instruction-following tasks where the user prefers a model with minimal inherent censorship.