Jiayi-Pan/Tiny-Vicuna-1B

Model Overview

Jiayi-Pan/Tiny-Vicuna-1B is a compact 1.1 billion parameter language model, developed by Jiayi-Pan. It is a fine-tuned variant of the TinyLlama base model, specifically trained on the WizardVicuna Dataset. This model is engineered to be fully compatible with the Vicuna-v1.5 series, making it a suitable option for developers familiar with that architecture.

Key Characteristics

• Base Model: Fine-tuned from TinyLlama (1.1B parameters).
• Training Data: Utilizes the WizardVicuna Dataset for instruction-following capabilities.
• Compatibility: Designed to be fully compatible with the Vicuna-v1.5 series.
• Efficiency: Its small parameter count makes it ideal for quick iterations and resource-constrained environments.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, Tiny-Vicuna-1B demonstrates foundational capabilities across various tasks. Its average score is 34.76, with specific metrics including:

• AI2 Reasoning Challenge (25-Shot): 33.45
• HellaSwag (10-Shot): 55.92
• MMLU (5-Shot): 25.45
• TruthfulQA (0-shot): 33.82
• Winogrande (5-shot): 58.41
• GSM8k (5-shot): 1.52

Use Cases

This model is particularly well-suited for:

• Rapid Prototyping: Its small size allows for fast training and inference cycles, accelerating experimental workflows.
• Educational Purposes: An accessible model for learning about fine-tuning and instruction-following LLMs.
• Resource-Limited Deployment: Suitable for applications where computational power or memory is a constraint.
• Early-Stage Development: Provides a solid base for initial explorations before scaling up to larger models.