Collections

Get structured data

These models generate structured data from natural language prompts. They enable exciting applications like:

  • Grammar-guided decoding: Constrain model outputs to valid grammar formats. Generate SQL queries, regular expressions, or other structured languages.
  • JSON schema validation: Ensure outputs conform to schema specifications. Describe your desired data structure in natural language and get back compliant JSON.
  • Code generation: Generate code that complies with syntactic and semantic constraints. Describe the program you want and get working code.
  • Dataframe parsing: Understand dataframe commands and produce properly formatted data. Manipulate and analyze dataframes using intuitive language.

Our Pick: andreasjansson/llama-2-13b-chat-gguf

For most users, we recommend the andreasjansson/llama-2-13b-chat-gguf model. It hits the sweet spot of popularity, speed, and general-purpose chat capabilities:

  • Based on the powerful Llama-2 model, it excels at understanding natural language intent
  • Grammar support allows it to generate structured data that matches your specifications
  • JSON schema validation ensures outputs are compliant with your desired format

Whether you need to parse natural language into SQL queries, generate JSON data matching a schema, or converse with data using a structured chat interface, this model is a great fit.

Runner Up: andreasjansson/codellama-7b-instruct-gguf

For specialized code generation tasks, andreasjansson/codellama-7b-instruct-gguf is an excellent alternative. While a bit slower than our top pick, it offers:

  • A powerful code generation model based on CodeLlama
  • Grammar support to constrain generated code to valid syntax
  • JSON schema validation to define expected inputs and outputs

If your structured data use case revolves heavily around generating code, this model is well worth considering. Its understanding of code semantics and syntax is top notch.

The other models in this collection are also quite capable for certain use cases. The larger Llama and CodeLlama variants offer even more power (at the cost of some speed). And andreasjansson/wizardcoder-python-34b-v1-gguf is great for Python code generation specifically. But for most users, our top recommendations offer the best balance of performance, flexibility, and ease of use.