llleb/mistral-7b-arc-qlora-exp7-3
The llleb/mistral-7b-arc-qlora-exp7-3 is a 7 billion parameter Mistral-7B-v0.1 model fine-tuned using QLoRA for the ARC-Challenge science multiple-choice question answering task. It was trained with a 4-bit NF4 QLoRA method and response-only loss on a dataset including ARC-Challenge, ARC-Easy, and OpenBookQA. This model is specifically optimized for scientific reasoning and factual recall in a multiple-choice format, demonstrating its strength in academic QA benchmarks.
Loading preview...
Model Overview
llleb/mistral-7b-arc-qlora-exp7-3 is a 7 billion parameter model based on mistralai/Mistral-7B-v0.1. It has been specifically fine-tuned using the QLoRA method to excel at the ARC-Challenge science multiple-choice question answering task. This model leverages 4-bit NF4 QLoRA with response-only loss, making it efficient for deployment while maintaining strong performance in its specialized domain.
Key Capabilities
- Specialized QA: Optimized for science multiple-choice question answering, particularly on the ARC-Challenge dataset.
- Efficient Fine-tuning: Utilizes QLoRA for efficient adaptation of the base Mistral-7B model.
- Targeted Training Data: Trained on a combination of ARC-Challenge, ARC-Easy (subset), and OpenBookQA datasets to enhance its scientific reasoning abilities.
Evaluation
The model's performance was evaluated using lm-evaluation-harness on the arc_challenge benchmark, employing a 25-shot prompting strategy. This evaluation confirms its proficiency in handling complex scientific questions in a multiple-choice format.
Good For
- Academic Question Answering: Ideal for tasks requiring factual recall and reasoning in scientific domains.
- Educational Tools: Can be integrated into systems for generating or answering science-related quizzes and tests.
- Research in QLoRA: Provides a practical example of QLoRA's effectiveness in task-specific fine-tuning for QA.