KohakuHub/docs/features/core/metadata.md

title, description, icon

title	description	icon
YAML Metadata	Comprehensive metadata system using YAML frontmatter in README.md files	i-carbon-information

YAML Metadata & Repository Cards

Comprehensive metadata system using YAML frontmatter in README.md files.

---

Overview

KohakuHub parses YAML frontmatter from README.md and displays it beautifully:

Metadata Tab shows:

• License with documentation links
• Languages (all languages, not just first 3)
• Framework & pipeline tags (models)
• Base models with clickable links
• Training datasets with clickable links
• Task categories (datasets)
• Size category (datasets)
• Evaluation metrics
• All other fields in individual cards

---

Supported Fields

All Repository Types

---
license: mit                    # License identifier
language:                       # Language codes
  - en
  - zh
tags:                          # General tags
  - computer-vision
  - pytorch
---

Models

---
library_name: transformers     # Framework
pipeline_tag: text-classification  # Task type
base_model:                    # Parent model(s)
  - bert-base-uncased
datasets:                      # Training datasets
  - glue
  - imdb
metrics:                       # Evaluation metrics
  - accuracy
  - f1
eval_results:                  # Structured evaluation
  - task: text-classification
    dataset: sst2
    metrics:
      accuracy: 0.953
      f1: 0.948
---

Datasets

---
task_categories:               # Task types
  - image-classification
  - text-to-image
size_categories: 1M<n<10M     # Dataset size
multilinguality: monolingual   # Language type
annotations_creators:          # How annotated
  - expert-generated
source_datasets:               # Source
  - original
---

---

Display System

Metadata Header (Top Bar)

Shows key info as badges:

• License (e.g., MIT License)
• Languages (up to 3, then "+N more")
• Framework (models)
• Pipeline tag (models)
• Size (datasets - prominent red badge)
• Task categories (up to 2, then "+N more")

Clicking "+N more" → Navigates to Metadata tab

Metadata Tab (Full Grid)

Organized as individual cards:

Models:

• License Card
• Languages Card (all languages)
• Framework Card (library + pipeline)
• Base Model Card (clickable links to models)
• Training Datasets Card (clickable links)
• Metrics Card (evaluation results table)

Datasets:

• License Card
• Languages Card
• Task Categories Card (all tasks as badges)
• Size Card (large badge)
• Multilinguality Card
• Annotations Card
• Source Card

All types:

• Additional fields shown as individual cards
• Arrays → Badge lists
• Objects → JSON code blocks
• Strings → Plain text

---

Tag Filtering

Metadata tags filtered out from "Tags" display:

Removed prefixes:

• dataset:*, license:*, region:*
• task_categories:*, size_categories:*
• format:*, modality:*
• diffusers:*, transformers:*
• endpoints_compatible, autotrain_compatible

Special handling:

• dataset:user/dataset-name → Moved to "Referenced Datasets" card
• Clean tags show: "art", "anime", "stable-diffusion" (meaningful tags only)

---

Referenced Datasets

Extract from tags:

Tags like dataset:KBlueLeaf/danbooru2023 become:

• Sidebar card: "Referenced Datasets"
• Clickable links to each dataset
• Expandable (shows 3, then "Show N more")

---

Examples

Complete Model Card

---
license: apache-2.0
language:
  - en
  - zh
library_name: transformers
pipeline_tag: text-classification
base_model: bert-base-uncased
datasets:
  - glue
  - imdb
metrics:
  - accuracy
tags:
  - sentiment-analysis
  - nlp
  - pytorch
---

# My Sentiment Classifier

This model classifies text sentiment...

Complete Dataset Card

---
license: cc-by-4.0
language: en
task_categories:
  - image-classification
  - text-to-image
size_categories: 1M<n<10M
multilinguality: monolingual
annotations_creators: expert-generated
source_datasets: original
tags:
  - art
  - anime
dataset:KBlueLeaf/danbooru2023-webp
---

# My Dataset

Contains 5M images...

---

Frontend Components

MarkdownViewer:

• Strips YAML frontmatter before rendering
• Parses frontmatter separately
• Content displayed without metadata block

MetadataHeader:

• Horizontal badges above tabs
• Shows only most important fields
• Clickable "+N more" badges

DetailedMetadataPanel:

• Grid of individual cards
• Each field in its own card
• Clean, organized layout

SidebarRelationshipsCard:

• Author always shown
• Base models (if any)
• Training datasets (if any)
• Expandable lists (shows 2, then more)

---

API Access

Get parsed metadata:

Metadata is returned in repo info responses:

# Get repo info (doesn't include metadata)
curl http://localhost:28080/api/models/username/repo

# Get tree (doesn't include metadata)
curl http://localhost:28080/api/models/username/repo/tree/main

# Download README to parse yourself
curl http://localhost:28080/models/username/repo/resolve/main/README.md

Client-side parsing:

• Frontend fetches README.md
• Parses YAML using js-yaml library
• Normalizes arrays (string|string[] → string[])
• Filters specialized vs general fields

---

See also: Repository Management, Web UI