Ollama 发布 LLaVA 1.6 视觉模型，新增 34B 版本
Vision models · Ollama Blog
Ollama 更新 LLaVA 视觉模型集合至 1.6 版本，支持最高 4 倍像素的图像分辨率，并在额外的文档、图表和示意图数据集上训练，提升了文字识别与推理能力。模型提供 7B、13B 和新增的 34B 三种参数规模，可通过 ollama run llava:7b、llava:13b、llava:34b 运行，采用 Apache 2.0 或 LLaMA 2 Community License 授权。CLI 中可直接引用 .jpg 或 .png 文件路径，Python、JavaScript 库和 REST API 则支持在 images 参数中传入 base64 编码文件。
正文
当前显示原文，中文全文译文尚未就绪。
February 2, 2024
New LLaVA models
The LLaVA (Large Language-and-Vision Assistant) model collection has been updated to version 1.6 supporting:
Higher image resolution: support for up to 4x more pixels, allowing the model to grasp more details; Improved text recognition and reasoning capabilities: trained on additional document, chart and diagram data sets; More permissive licenses: distributed via the Apache 2.0 license or the LLaMA 2 Community License
These models are available in three parameter sizes. 7B, 13B and a new 34B model:
ollama run llava:7b; ollama run llava:13b; ollama run llava:34b
Usage
CLI
To use a vision model with ollama run, reference .jpg or .png files using file paths:
Python
JavaScript
Note: in the Ollama Python and JavaScript libraries and the REST API, base64-encoded files can be provided in the images parameter. See the full API docs for more examples on providing images to vision models.
The LLaVA 1.6 model family is published by an incredible team. Information on how the models were trained is available on the LLaVA website, as well as benchmark results comparing LLaVA 1.6 to leading open-source and proprietary models.