Model Overview

The andrijdavid/tinyllama-dare is a 1.1 billion parameter language model developed by andrijdavid. It is a merged model, combining the strengths of five different TinyLlama-based models: aihub-app/zyte-1B, BEE-spoke-data/TinyLlama-3T-1.1bee, sreeramajay/TinyLlama-1.1B-orca-v1.0, vihangd/DopeyTinyLlama-1.1B-v1, and kevin009/lamatama. This merge aims to consolidate diverse capabilities into a single, compact model with a 2048-token context length.

Performance Highlights

Evaluated on the Open LLM Leaderboard, the model achieved an average score of 38.64. Key benchmark results include:

• HellaSwag (10-Shot): 62.78
• Winogrande (5-shot): 65.90
• AI2 Reasoning Challenge (25-Shot): 37.29
• TruthfulQA (0-shot): 39.01
• MMLU (5-Shot): 25.20
• GSM8k (5-shot): 1.67

Considerations for Use

Users should be aware of common limitations inherent in language models, including potential for factual inaccuracies, biases from training data, and hallucinations. The model is provided "as is," and users are advised to verify outputs, especially for critical applications.