NewstaR/Starlight-13B

Starlight-13B: A Conversational Transformer Model

NewstaR/Starlight-13B is a 13 billion parameter transformer model developed by NewstaR, specifically trained for conversational text generation. It utilizes the Alpaca instruction template, making it suitable for various dialogue-based applications. The model was trained on a combination of the AverageData and Above the Clouds datasets, contributing to its language modeling capabilities.

Key Capabilities & Performance

Starlight-13B demonstrates solid performance across several benchmarks, as evaluated on the Open LLM Leaderboard:

• Average Score: 58.63
• HellaSwag (10-shot): 82.15
• MMLU (5-shot): 55.67
• ARC (25-shot): 59.3
• TruthfulQA (0-shot): 37.39

These scores indicate its proficiency in common language understanding and generation tasks. Detailed benchmark results are available on the Open LLM Leaderboard.

Intended Use Cases

Starlight-13B is primarily intended for:

• Chatbots: Generating conversational responses in interactive systems.
• Content Creation: Assisting with the generation of various forms of text content.

Limitations and Safe Use

Users should be aware that Starlight-13B, like many large language models, may:

• Hallucinate or generate incorrect information.
• Exhibit biases or toxicity present in its training data.
• Require significant computational resources due to its size.

It is crucial to monitor outputs, implement safeguards, and avoid using the model for high-stakes or safety-critical applications.