What would you like to see in the OAIB specification?

[kisamoto]

The first step of defining this specification is to collect key information that should be useful to consumers but easy to implement for AI service providers.

We welcome feedback and ideas for additional features. Keep in mind that this is purely the initial list, the formal specification will then be written to encapsulate them all and will enter a new phase for discussion.

Terminology

Try to define some base terminology:

Term	Description
OAIB	The "Open AI Broker" or "The Open Source AI Broker"
Provider	A service that provides AI services such as models (either for download or hosted) or Agents (either for download or hosted)
Model	A program trained on a dataset
Agent	At this stage an Agent is a custom prompt or a fine-tuned model however it may expand to also include the ability to take additional actions (e.g. call end-points)

Initial list of attributes

We need some basic information for both models and Agents to be returned in the catalog:

• OAIB version that the provider is implementing
• Provider information
  • Identifier
  • Name
  • URL for more information
  • Contact details (optional)
  • Legal
    • Terms of service
    • Privacy policy
• Agent/model information
  • Identifier
  • Name
  • Slug (can be used as a version of the name in URLs)
  • Description
  • License
  • Creator information (will fall back to provider information if not provided)
    • Name
    • Contact details (optional)
  • URL for more information (e.g. HuggingFace)
  • Metadata (optional - freeform object of key-value pairs of strings)
  • Date created
  • Date update
  • Date added to catalog
  • Tags (array of strings)
    • Will be predefined in the spec
  • Pricing
    • Free? true/false
    • Plans (if not free, array of plans)
      • Name
      • Currency
      • Amount
      • Unit
    • Metadata (additional pricing specific metadata)
• Model specific information
  • Variant (e.g. instruct/base)
  • Number of parameters
  • Context length
  • Link to more details (e.g. HuggingFace)
  • Quantization length (e.g. Q5_K_M)
  • Distribution format (e.g. GGUF)
• Agent specific information
  • Prompt (optional as it may be kept private on some marketplaces)
  • Number of tokens in prompt
  • Recommended model ID (model identifier must be present in the model list)
  • Parameters (optional)
    • A list of objects
      • Name
      • Description
      • Example list
  • Response (object)
    • Type (e.g. JSON/text/code)
    • Example list
  • Usage examples (list of lists, can be used also as prompts to other models)
    • Each entry is a list of 1 or more examples

[kisamoto]

Hardware requirements for models may be useful however they may also be hard to quantify on an even field and would only be relevant if the consumer was running the model on their own hardware.

(V)RAM, possible CPU/GPU requirements etc.
- e.g memory: 8GB and 20 TOPS