Base Model

• Foundation: Derived from llava:13b, with a series of other models built as well

Purpose

• This model specializes in generating high-quality alt text for social media posts, particularly focusing on photographs and memes.
• The descriptions are designed to be concise, accessible, and ready for immediate use in alt text fields to support screen reader compatibility.

Key Features

1. Optimized for Accessibility

• Generates precise and descriptive alt text to enhance inclusivity for visually impaired users on social media platforms.
• Ensures seamless integration with screen readers through direct copy-paste functionality.

2. Content-Specific Guidelines

• Photographs: Describes visible elements with attention to composition, layout, and notable details.
• Memes: Captures both visual and textual elements, providing necessary context.
• Text Transcription: Includes exact replication of visible text for accuracy.

3. Ethical and Efficient AI Use

• Incorporates adult content processing with clear, professional descriptions to ensure comprehensive coverage of diverse content types.
• Adheres to strict guidelines to maintain clarity and conciseness under 1000 characters.

4. Customizable Parameters

• Temperature: Set at 0.1 for controlled and focused outputs.
• Predict Length: Configured for up to 1024 tokens to accommodate detailed descriptions.

5. License and Attribution

• Licensed under a derivative of the LLAMA 2 Community License Agreement.
• Custom modifications include integration with the Assisted Space platform for workflow enhancement and ethical AI practices.
• Attribution provided at Luke Steuber’s website, with additional details available on the Assisted Space Hub.

Prompt Template

The model uses a structured prompt format:

{{ .System }} USER: {{ .Prompt }} ASSISTANT:

SYSTEM Instructions:

• Define the role as an Alt Text Specialist.
• Outline clear rules for image description, emphasizing specificity, clarity, and accessibility.
• Ensure readiness for social media alt fields without requiring further edits.