Phi 3.5 Vision Instruct is an open-source language model by microsoft. Features: 4.1b LLM, VRAM: 8.3GB, Context: 128K, License: mit, Instruction-Based, LLM Explorer Score: 0.27.
multimodal model, text generation, vision understanding
Use Cases
Areas:
broad commercial, research use, English
Applications:
general purpose AI systems with visual and text input, Memory/compute constrained environments, Latency bound scenarios, General image understanding, Optical character recognition, Chart and table understanding, Multi-image comparison, Multi-image or video clip summarization
Primary Use Cases:
research on language and multimodal models, building block for generative AI powered features
Limitations:
models are not specifically designed or evaluated for all downstream purposes
Considerations:
Developers should consider common limitations of language models and ensure use complies with applicable laws.
Additional Notes
Developers are responsible for mitigating biases and ensuring accuracy and safety in their use cases.
Supported Languages
multilingual (primarily focused on English text)
Training Details
Data Sources:
synthetic data, filtered publicly available websites, high-quality educational data, image-text interleave data, synthetic 'textbook-like' data for teaching math, coding, reasoning, etc., created multi-image and video data
Data Volume:
500B tokens
Methodology:
supervised fine-tuning and direct preference optimization
Context Length:
128000
Training Time:
6 days
Hardware Used:
256 A100-80G GPUs
Model Architecture:
includes image encoder, connector, projector, and Phi-3 Mini language model
Safety Evaluation
Methodologies:
red teaming, adversarial conversation simulations, safety evaluation benchmark datasets
Risk Categories:
production of undesirable outputs across multiple risk categories
Ethical Considerations:
Leveraged human-labeled and synthetic datasets focusing on safety categories
Responsible Ai Considerations
Fairness:
Limitations may still be present due to differing levels of representation of different groups or societal biases.
Transparency:
Developers should inform end-users that they are interacting with an AI system.
Accountability:
Developers are responsible for ensuring compliance with relevant laws and regulations.
Mitigation Strategies:
Apply responsible AI best practices, use safety classifiers or custom solutions, ensure transparency and accurate information.
Input Output
Input Format:
Best suited for prompts using the chat format.
Accepted Modalities:
Text, Image
Output Format:
Generated text in response to input.
Performance Tips:
Set num_crops=4 for multi-frame and num_crops=16 for single-frame for best performance.
Release Notes
Date:
August 2024
Notes:
Model enables multi-frame image understanding, improved single image benchmark performance, supports wider range of applications.