# SelfJev SelfJev is a self-hosted 4B decision model for turning text, images, and questions into typed answers with probabilities. It helps developers run structured classification, routing, review, and policy workflows on infrastructure they control. - Findings24: - Site: - Preços: Freemium - Categorias: Infraestrutura para agentes de IA, Plataformas de inferência de IA, Reconhecimento de Imagem IA ## O que é SelfJev? SelfJev is a self-hosted 4B decision model for applications that need structured choices rather than generated prose. It accepts text, images, or both with a set of questions, then returns typed answers and probabilities. Supported answer formats include yes/no, pick one, rate a result, and select all that apply. Its shared-prefix inference engine reads the common input once and reuses that work across multiple questions. This makes it suitable for extracting several decisions from one support message, document, AI response, or image. The model runs as an HTTP service on infrastructure controlled by the user, with a Python client available for application integration. The default release supports vision and text through SelfJev’s native engine. SelfJev also provides fine-tuning tools, public model weights, evaluation datasets, scoring tools, and recorded experiment reports. Published scores are limited to the project’s test suites and should be treated as evidence for evaluation, not a guarantee of performance on new data. ## O que SelfJev pode fazer? ### Typed decision outputs Returns structured answers for yes/no questions, single-choice selection, ratings, and multi-select questions, along with probabilities that application code can use. ### Shared-prefix inference Reads a shared text or image input once and branches several questions from that work, supporting multi-question requests without separately processing the same context for every question. ### Vision and text input The default SelfJev-4B release accepts images and text through the native tree engine. Images can be provided as paths, bytes, PIL images, or base64 data URLs over HTTP. ### Self-hosted HTTP service Runs on a user-controlled Linux machine with an NVIDIA GPU and exposes an API for applications. The documented setup includes health checks, API-key configuration, and a lightweight client that does not require a GPU. ### Fine-tuning for domain decisions Supervised fine-tuning tools let users train lightweight adapters on verified examples from their own workflow, including image-based training rows, before evaluating and serving the result. ### Open evaluation record The project publishes model releases, Decision Bench data, scoring tools, methodology, and recorded experiments so users can inspect how reported results were produced. ## Casos de uso ### Support ticket routing Ask whether a customer needs a refund, which team should handle the request, and which topics apply, then pass the resulting fields into routing or workflow automation. ### AI response review Evaluate an AI-generated answer for quality, accuracy, supporting evidence, safety, or attempts to bypass instructions before it reaches a user or downstream system. ### Policy and guardrail checks Turn a policy into explicit questions and use the returned choices and probabilities to filter requests, apply controls, or flag cases for human review. ### Image-based classification Ask several questions about a photo, such as category or condition, using the same image as shared context. The site documents examples involving pets, clothing, plants, rice, satellite land use, and waste. ### Workflow-specific decision models Fine-tune an adapter on verified examples containing a team’s labels, policies, vocabulary, or edge cases, then compare results on held-out examples before deployment. ## Perguntas frequentes ### What does SelfJev return? It returns typed answers and probabilities. The documented answer types are yes/no, pick one, rate a result, and select all that apply; it does not generate a conversational text response. ### What hardware is needed to run SelfJev? The quickstart calls for a Linux machine with an NVIDIA GPU and recommends starting with 24 GB of VRAM, 4 vCPUs, 16–32 GB of system RAM, and 50 GB of free disk. These figures are planning recommendations, not a tested minimum. ### Can SelfJev process images? Yes. The default vision-and-text release accepts images through the native engine. The Python client accepts a path, image bytes, or a PIL image; HTTP requests can use a base64 data URL or text and image parts. ### How does an application call the model? SelfJev runs as an HTTP server, and applications can use the lightweight Python client. The quickstart also documents pointing an existing TypeSafe Python SDK installation at the SelfJev server by changing its base URL and API key settings. ### Can the model be adapted to a custom workflow? Yes. The site documents supervised fine-tuning with verified examples from a workflow, including image training rows. Users are advised to evaluate on held-out examples before serving the adapted model. ## Informações rápidas - **Product type**: Self-hosted 4B decision model - **Primary interface**: HTTP API with Python client - **Inputs**: Text, images, or text and images - **Outputs**: Typed answers with probabilities; no generated text - **Default runtime**: Native tree engine on an NVIDIA GPU - **Published evaluation**: 96.1% expected-answer match on the project’s 1,991-question text-decisions test for the default vision-and-text model