Dolly v2 (7B) API

Basic Information

• Model Type: Instruction-following Large Language Model

Overview:Dolly v2 (7B) is an instruction-following large language model trained on the Databricks machine learning platform, licensed for commercial use. It is based on the Pythia-6.9b model and fine-tuned on a dataset of approximately 15,000 instruction/response pairs.Key Features:

• Instruction-following capabilities
• Open-source and commercially licensed
• Fine-tuned on high-quality training data
• Relatively small model size (6.9 billion parameters)

Intended Use:Dolly v2 (7B) is designed for various natural language processing tasks, including:

Technical Details

Architecture:Dolly v2 (7B) is based on the Pythia-6.9b architecture, which is a transformer-based model.Training Data:

• Data Source and Size: The model was fine-tuned on the databricks-dolly-15k dataset, containing approximately 15,000 instruction/response pairs generated by Databricks employees.
• Knowledge Cutoff: The model's knowledge is based on the Pythia-6.9b pre-training, with additional instruction-following capabilities from the fine-tuning process.

Performance Metrics:While Dolly v2 (7B) is not state-of-the-art, it demonstrates surprisingly high-quality instruction-following behavior. Some benchmark results include:

BenchmarkScoreARC (25-shot) 0.392 HellaSwag (10-shot) 0.633838 MMLU (5-shot) 0.406997 TruthfulQA (0-shot)0.444444

Ethical Considerations

Dolly v2 (7B) was developed with ethical considerations in mind. The training data does not contain obscenity, intellectual property, or personally identifying information about non-public figures. However, it may reflect biases present in the data generated by Databricks employees.

Licensing

Dolly v2 (7B) is released under the Apache 2.0 license, which allows for both research and commercial use.

Limitations

• Not designed to perform competitively with more modern model architectures
• Struggles with syntactically complex prompts
• Limited capabilities in programming problems and mathematical operations
• May contain factual errors and exhibit hallucination
• Reflects potential biases of Databricks employees in the training data

Use cases

• Chatbots and conversational assistants
• Content generation and summarization
• Educational tools and tutoring
• Text classification and information extraction
• Research and prototyping in NLP
• Commercial and open-source AI applications

Capabilities

• Function calling
• Streaming
• Structured outputs