Skip to content

Latest commit

 

History

History
104 lines (79 loc) · 4.18 KB

README.md

File metadata and controls

104 lines (79 loc) · 4.18 KB

LLM Metadata Repository

This repository provides structured, machine-readable metadata for a wide range of large language models (LLMs). It is designed to support tools and applications that require detailed information about model capabilities, configurations, and supported parameters. Used by BasiliskLLM and OpenAI NVDA Add-on.

This metadata enables:

  • Dynamic population of model selection UIs
  • Feature-aware prompting and parameter tuning
  • Compatibility and capability checks for downstream tools

The metadata is stored in JSON format and is inspired by the OpenRouter API model listing schema.

Each JSON file in the data/ directory contains a list of model objects with the following structure:

{
  "id": "gpt-5",
  "name": "GPT-5",
  "description": "OpenAI’s most advanced model...",
  "created": 1754587413,
  "context_length": 400000,
  "architecture": {
    "modality": "text+image->text",
    "input_modalities": ["text", "image", "file"],
    "output_modalities": ["text"],
    "tokenizer": "GPT",
    "instruct_type": null
  },
  "top_provider": {
    "context_length": 400000,
    "max_completion_tokens": 128000,
    "is_moderated": true
  },
  "supported_parameters": [
    "max_tokens",
    "temperature",
    "response_format",
    "structured_outputs"
  ]
}

Metadata Fields

Top-Level Fields

  • id: Unique model identifier (e.g., gpt-4-turbo)
  • name: Human-readable model name
  • description: Summary of model capabilities and use cases
  • created: Unix timestamp of model release (used for sorting by release date; models with -latest in the id are listed first)
  • context_length: Maximum input context length (in tokens)

architecture

  • modality: Overall input/output modality (e.g., text->text, text+image->text)
  • input_modalities: Supported input types (e.g., text, image, file)
  • output_modalities: Supported output types (e.g., text)
  • tokenizer: Tokenizer type used by the model (e.g., GPT)
  • instruct_type: Instruction tuning format (e.g., chatml, alpaca), or null if not applicable

top_provider

  • context_length: Maximum context length supported by the top provider
  • max_completion_tokens: Maximum number of tokens in a single response
  • is_moderated: Indicates whether the model is subject to content moderation

supported_parameters

A list of tunable parameters supported by the model, such as:

  • max_tokens
  • temperature
  • top_p
  • frequency_penalty
  • presence_penalty
  • tools
  • seed
  • response_format
  • structured_outputs

Data Sources

Metadata is curated from official provider documentation and APIs. Official model listings:

Data may also be synchronized from OpenRouter API which aggregates models from multiple providers.

Contributing

Contributions are welcome! To add or update metadata for a model:

  1. Fork the repository
  2. Add or edit the appropriate JSON file in the data/ directory
  3. Submit a pull request

Please ensure your JSON is valid and follows the schema outlined above. Prefer official provider documentation when updating model IDs, descriptions, or parameters.

Models are sorted with *-latest aliases first, then by release date (newest first). Run python sort_models.py after editing to maintain sort order.