# Valen

> Valen is a multimodal System One decision model. Text, images and video in; structured probabilities over the candidates you supply out.

Every page has a Markdown version: append `.md` to a docs or cookbook path.

## API

- [OpenAPI spec](https://valen-team.github.io/openapi.json): request and response schemas for System One, Decisions and List Models
- [Agent skill](https://github.com/Valen-Team/skills): integration craft for coding agents

## Docs

- [Documentation](https://valen-team.github.io/docs.md): Provide text, images or video, define a question, and get probabilities over your candidates.
- [Build with us](https://valen-team.github.io/docs/contributing.md): We welcome contributions to models, evaluation tasks, data, documentation, and agent applications.
- [JevalKit](https://valen-team.github.io/docs/jevalkit.md): A toolkit for evaluating Jev-like decision models.
- [Benchmark protocols](https://valen-team.github.io/docs/jevalkit/benchmarks.md): Project overview · Running evaluations · Model integrations
- [Model integrations](https://valen-team.github.io/docs/jevalkit/models.md): Project overview · Running evaluations · Scoring protocols
- [Run your first evaluation](https://valen-team.github.io/docs/jevalkit/quickstart.md): Project overview · Model integrations · Scoring protocols
- [Models, datasets, and demos](https://valen-team.github.io/docs/resources.md): Model weights
- [Valen](https://valen-team.github.io/docs/valen.md): An open multimodal System One decision model.
- [Build with a coding agent](https://valen-team.github.io/docs/valen/agent-skill.md): Install the Valen skill so Claude Code, Cursor, Codex and other coding agents design Valen integrations well.
- [API calls](https://valen-team.github.io/docs/valen/api.md): Hosted APIs and local deployments support two protocols:
- [API data handling and retention](https://valen-team.github.io/docs/valen/api/data-handling.md): This page describes the hosted Valen API. For a self-hosted deployment, data handling depends on the operator's configuration.
- [Decisions](https://valen-team.github.io/docs/valen/api/decisions.md): Send text or images to the hosted API and ask predicate, choice, or score questions. The API returns results and associated probabilities for each question. For video, extract frames first and submit them as image inputs.
- [List Models](https://valen-team.github.io/docs/valen/api/list-models.md): List models exposed by the hosted service. Use a returned id as the model field in inference requests.
- [System One](https://valen-team.github.io/docs/valen/api/systemone.md): Send text or images to the hosted API and ask predicate, choice, or score questions. The API returns results and associated probabilities for each question. For video, extract frames first and submit them as image inputs.
- [Released model architecture](https://valen-team.github.io/docs/valen/architecture.md): Valen uses Qwen3.5 to understand the input text, images or video. Mixer scores each candidate answer, and softmax converts those scores to probabilities. Several questions about the same input can share one backbone computation.
- [Calibrate and automate](https://valen-team.github.io/docs/valen/calibration.md): Use measured results to choose what Valen routes automatically and what needs review.
- [Training checkpoint inference](https://valen-team.github.io/docs/valen/checkpoint-inference.md): This page covers per-question inference and evaluation in the valen package. Full-game Sokoban evaluation uses the separate evaluation.sokoban command; see the task evaluation guide for data generation, policy comparison and trajectory replay.
- [Save, resume and export](https://valen-team.github.io/docs/valen/checkpoints.md): Training writes logs and configuration to the output directory and overwrites latest/ on each checkpoint save. Copy the complete directory after saving finishes to retain a step. Frozen weights still depend on the original model_path.
- [Choose a version](https://valen-team.github.io/docs/valen/choosing.md): Choose a path for your device and task first, then choose a model size. To try Valen without downloading a model, use the web experience or free API.
- [Training configuration](https://valen-team.github.io/docs/valen/configuration.md): Copy an existing recipe instead of starting from an empty JSON:
- [Data format](https://valen-team.github.io/docs/valen/data-format.md): Training and evaluation datasets use UTF-8 JSONL: one JSON record per line, with blank lines ignored. The reader loads the whole file into memory at once, and an empty file raises an error. The examples below are expanded for readability; collapse each record onto a single line when writing .jsonl. Ready-to-use files are in data/smoke.
- [Jev Decision Index](https://valen-team.github.io/docs/valen/decision-index.md): Jev Decision Index compares models on decision tasks. Valen-4B scores 37.57 overall and 38.54 on the public suite; results for all three sizes are below.
- [Evaluation](https://valen-team.github.io/docs/valen/evaluation.md): Use the same independent test before and after fine-tuning. Keep candidates, order, temperature, media budgets and backend fixed. Select settings on validation before inspecting test results.
- [First call](https://valen-team.github.io/docs/valen/getting-started.md): Valen scores supplied candidates for text, images and video. Start with the hosted text API; availability and quotas depend on the service. No local model or training environment is needed.
- [GGUF local service](https://valen-team.github.io/docs/valen/gguf.md): These models require the Valen-patched llama.cpp/Ollama. Support in official unmodified releases is not claimed. For the quickest online trial, see version selection.
- [Known limitations](https://valen-team.github.io/docs/valen/known-limitations.md): Where Valen's answers become unreliable, measured on public benchmarks, and how to restructure each case.
- [ONNX inference and SDKs](https://valen-team.github.io/docs/valen/onnx.md): Run Valen with Python, Node.js, browser WebGPU or a native C++ text integration.
- [Use probabilities in applications](https://valen-team.github.io/docs/valen/probabilities.md): Choice distributes probability over the supplied mutually exclusive candidates and selects the maximum. Predicate / Noul gives P(true) for one statement; separate predicates are not mutually exclusive. Score distributes probability over ordered grade descriptions and returns sum(ip[i]), with zero-based indices. Labels do not create numeric business values: low/medium/high need not mean equally spaced cost. Use your own explicit utility mapping sum(utility[i]p[i]) when appropriate; validate that mapping separately.
- [Write questions that work](https://valen-team.github.io/docs/valen/question-design.md): How to turn a decision into Valen questions, design text and image state, write candidates, and restructure around recorded hard cases.
- [Transformers local inference](https://valen-team.github.io/docs/valen/quickstart.md): Install the dependencies and run Valen with either inference protocol.
- [Multi-question inference](https://valen-team.github.io/docs/valen/qwen-shared-state.md): Ask several questions about the same text, image or video. Valen puts them in one shared sequence and scores their candidates with one backbone forward pass.
- [Released-model evidence](https://valen-team.github.io/docs/valen/reliability.md): Compare calibration, automation coverage and error rates for released Valen models.
- [SGLang image deployment](https://valen-team.github.io/docs/valen/sglang-image.md): Run a local HTTP service with a Docker image containing the SGLang runtime and Valen plugin.
- [SGLang plugin deployment](https://valen-team.github.io/docs/valen/sglang-plugin.md): Install the Valen plugin in an official SGLang environment and start a local HTTP inference service.
- [SGLang deployment](https://valen-team.github.io/docs/valen/sglang.md): Choose the deployment route that fits your environment:
- [Fine-tuning](https://valen-team.github.io/docs/valen/training.md): Fine-tuning uses your labeled examples to adapt an existing model—for example, to your company’s support-ticket categories. Start with Valen-0.8B and LoRA, which updates a small subset of parameters for an initial experiment.
- [vLLM image deployment](https://valen-team.github.io/docs/valen/vllm-image.md): Run a local HTTP service with a Docker image containing the vLLM runtime and Valen plugin.
- [vLLM plugin deployment](https://valen-team.github.io/docs/valen/vllm-plugin.md): Install the Valen plugin in an official vLLM environment and start a local HTTP inference service.
- [vLLM deployment](https://valen-team.github.io/docs/valen/vllm.md): Choose the deployment route that fits your environment:

## Cookbook

- [Cookbook](https://valen-team.github.io/cookbook.md): Tested Valen recipes from input to business output, grouped by the goal they serve.
- [API basics](https://valen-team.github.io/cookbook/api-quickstart.md): Run the original text, image and sampled video-frame examples.
- [Calibrate ticket routing](https://valen-team.github.io/cookbook/calibrated-routing.md): Fit a reusable routing rule on your own tickets and send the rest for review.
- [Check claims against their citations](https://valen-team.github.io/cookbook/citation-check.md): Two predicates per claim separate supported, contradicted and unsupported claims; the ones the source does not settle go to an editor.
- [Score photos once, weigh them per storefront](https://valen-team.github.io/cookbook/composite-scoring.md): Score listing photos on lighting, background and framing once; each storefront ranks them with its own weights in code.
- [Check scanned forms against the record](https://valen-team.github.io/cookbook/document-check.md): Ask one question per field of a scanned loan form, then accept, reject or send the scan for review in code.
- [Domain fine-tuning](https://valen-team.github.io/cookbook/domain-finetuning.md): Classify requests as address changes, cancellations or other requests. This small example covers data checks, LoRA training, resume and before/after evaluation.
- [Find when events happen in a clip](https://valen-team.github.io/cookbook/event-monitoring.md): Ask the same predicates of evenly spaced frames and smooth the probabilities into event times in code.
- [Turn a command into a function call](https://valen-team.github.io/cookbook/function-calling.md): One Valen request picks the function and fills every closed-set argument; code reads only the chosen function's answers.
- [Guardrails for LLM replies](https://valen-team.github.io/cookbook/llm-guardrails.md): One predicate per policy rule decides whether an LLM's draft reply is sent, blocked, reviewed or regenerated.
- [Local inference basics](https://valen-team.github.io/cookbook/local-inference.md): Use both protocols with offline native or ONNX models.
- [Migration map](https://valen-team.github.io/cookbook/migration.md): The four notebooks are two tutorials in two languages, not four distinct applications. This inventory separates duplicated material from content worth preserving. Complete tutorial code and JSON samples live only in the website repository; API and deployment references retain their own responsibilities.
- [Screen product images](https://valen-team.github.io/cookbook/product-screening.md): Check photographic print listings for image/category mismatch.
- [Rerank retrieved passages](https://valen-team.github.io/cookbook/rerank.md): Ask whether each retrieved help-center passage answers the customer's question, sort by probability, and escalate when none does.
- [Build a support router with the agent skill](https://valen-team.github.io/cookbook/skill-support-router.md): A coding agent with the Valen skill turned a one-page brief into a support router that scores 98% on 60 unseen messages.
- [Route support tickets](https://valen-team.github.io/cookbook/ticket-routing.md): Produce a queue, review decision and priority from a customer message.
- [When no option fits: handling unknown categories](https://valen-team.github.io/cookbook/unknown-categories.md): Compare explicit none choices and review thresholds on Valen support-ticket decisions, with reproducible predictions and failure cases.
- [Validation, performance and versions](https://valen-team.github.io/cookbook/validation.md): Validated on 26 October 2026. This page reads the version manifest, performance rows and capability matrix directly from their machine-readable sources. The benchmark uses real 0.8B model weights with a small fixed text case and synthetic diagnostic image/video inputs. These are not representative task-accuracy or production-load benchmarks.
- [Pick the right date from an email](https://valen-team.github.io/cookbook/value-extraction.md): Code finds every date in a supplier email; Valen picks the new delivery date, and a predicate catches emails that have none.
- [Tag short videos](https://valen-team.github.io/cookbook/video-tagging.md): Produce a visual attribute tag with frame provenance and uncertainty handling.

## Optional

- [All pages in one file](https://valen-team.github.io/llms-full.txt)

简体中文版本：https://valen-team.github.io/zh-Hans/llms.txt