Add SEO descriptions and metadata across docs pages

deployment/gpu-inference/baseten.mdx

@@ -1,10 +1,10 @@
 ---
 title: "Baseten"
-description: "Baseten is an AI infrastructure platform for deploying and serving ML models with optimized inference, autoscaling, and multi-cloud support."
+description: "Deploy Liquid AI LFM models on Baseten with optimized inference, autoscaling, and multi-cloud support."
 ---
 
 <Tip>
   Use Baseten for production ML model deployments with optimized inference, autoscaling, and multi-cloud support.
 </Tip>
 
 ## Clone the repository[​](#clone-the-repository "Direct link to Clone the repository")
@@ -15,7 +15,7 @@
 
 ## Deployment[​](#deployment "Direct link to Deployment")
 
 The deployment script is based on Baseten's [documentation](https://docs.baseten.co/examples/vllm) of `Run any LLM with vLLM`.
 
 Launch command:
 
@@ -42,5 +42,5 @@
 ```
 
 <Note>
   Baseten endpoints expect the `Api-Key` prefix in the `Authorization` header.
 </Note>

deployment/gpu-inference/fal.mdx

@@ -1,10 +1,10 @@
 ---
 title: "Fal"
-description: "Fal is a serverless generative media platform offering lightning-fast inference for AI models for image, video, and audio generation."
+description: "Deploy Liquid AI LFM models on Fal for serverless inference with fast autoscaling."
 ---
 
 <Tip>
   Use Fal for serverless cloud deployments with lightning-fast inference, autoscaling, and easy API access.
 </Tip>
 
 ## Clone the repository
@@ -56,5 +56,5 @@
 ```
 
 <Note>
   Fal endpoints expect the `Key` prefix in the `Authorization` header.
 </Note>

deployment/gpu-inference/modal.mdx

@@ -1,10 +1,10 @@
 ---
 title: "Modal"
-description: "Modal is a serverless cloud platform for running AI/ML workloads with instant autoscaling on GPUs and CPUs."
+description: "Deploy Liquid AI LFM models on Modal with serverless GPU inference and instant autoscaling."
 ---
 
 <Tip>
   Use Modal for serverless cloud deployments with instant autoscaling, GPU access, and production-ready inference serving.
 </Tip>
 
 ## Clone the repository
@@ -31,7 +31,7 @@
 
 ## Production deployment
 
 * Since vLLM takes over 2 min to cold start, if you run the inference server for production, it is recommended to keep a minimum number of warm instances with `min_containers = 1` and `buffer_containers = 1`. The `buffer_containers` config is necessary because all Modal GPUs are subject to [preemption](https://modal.com/docs/guide/preemption). See [docs](https://modal.com/docs/guide/cold-start#overprovision-resources-with-min_containers-and-buffer_containers) for details about cold start performance tuning.
 * Warm up the vLLM server after deployment by sending a single request. The warm-up process is included in the [deploy-vllm.py](https://github.com/Liquid4All/lfm-inference/blob/main/modal/deploy-vllm.py) script already.
 
 ## Test commands

deployment/gpu-inference/transformers.mdx

@@ -1,6 +1,6 @@
 ---
 title: "Transformers"
-description: "Transformers is a library for inference and training of pretrained models."
+description: "Run Liquid AI LFM models with Hugging Face Transformers. Load, configure, and generate with LFM2.5 models in Python."
 ---
 
 <Tip>
@@ -11,7 +11,7 @@
 
 <div className="colab-link">
   <a href="https://colab.research.google.com/github/Liquid4All/docs/blob/main/notebooks/LFM2_Inference_with_Transformers.ipynb" target="_blank">
     <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open in Colab" />
   </a>
 </div>
 
@@ -70,10 +70,10 @@
 
 * **`model_id`**: Can be a Hugging Face model ID (e.g., `"LiquidAI/LFM2.5-1.2B-Instruct"`) or a local path
 * **`device_map="auto"`**: Automatically distributes across available GPUs/CPU (requires `accelerate`). Use `device="cuda"` for single GPU or `device="cpu"` for CPU only
 * **`dtype="bfloat16"`**: Recommended for modern GPUs. Use `"auto"` for automatic selection, or `"float32"` (slower, more memory)
 
 <Accordion title="Click to see a pipeline() example">
   The [`pipeline()`](https://huggingface.co/docs/transformers/en/main_classes/pipelines) interface provides a simpler API for text generation with automatic chat template handling. It wraps model loading and tokenization, making it ideal for quick prototyping.
 
   ```python
   from transformers import pipeline
@@ -98,10 +98,10 @@
 
   * **`"text-generation"`**: Task type for the pipeline
   * **`model_name_or_path`**: Model ID (e.g., `"LiquidAI/LFM2.5-1.2B-Instruct"`) or local path (download locally with `hf download --local-dir ./LFM2.5-1.2B-Instruct LiquidAI/LFM2.5-1.2B-Instruct`)
   * **`dtype="auto"`**: Automatically selects optimal dtype (`bfloat16` on modern devices). Can use `"bfloat16"` explicitly or `"float32"` (slower, more memory)
   * **`device_map="auto"`**: Automatically distributes across available GPUs/CPU (requires `accelerate`). Alternative: `device="cuda"` for single GPU, `device="cpu"` for CPU only. Don't mix `device_map` and `device`
 
   The pipeline automatically handles chat templates and tokenization, returning structured output with the generated text.
 </Accordion>
 
 ### Generation Parameters
@@ -110,7 +110,7 @@
 
 * **`do_sample`** (`bool`): Enable sampling (`True`) or greedy decoding (`False`, default)
 * **`temperature`** (`float`, default 1.0): Controls randomness (0.0 = deterministic, higher = more random). Typical range: 0.1-2.0
 * **`top_p`** (`float`, default 1.0): Nucleus sampling - limits to tokens with cumulative probability ≤ top\_p. Typical range: 0.1-1.0
 * **`top_k`** (`int`, default 50): Limits to top-k most probable tokens. Typical range: 1-100
 * **`min_p`** (`float`): Minimum token probability threshold. Typical range: 0.01-0.2
 * **`max_new_tokens`** (`int`): Maximum number of tokens to generate (preferred over `max_length`)

deployment/on-device/atomic-chat.mdx

@@ -1,6 +1,6 @@
 ---
 title: "Atomic Chat"
-description: "Atomic Chat is a desktop and mobile app for running LLMs locally with a graphical user interface."
+description: "Run Liquid AI LFM models locally with Atomic Chat, a desktop and mobile app with a graphical interface."
 ---
 
 <Tip>
@@ -42,7 +42,7 @@
 Control text generation behavior using the GUI sidebar or API parameters. Key parameters:
 
 * **`temperature`** (`float`, default 1.0): Controls randomness (0.0 = deterministic, higher = more random). Typical range: 0.1-2.0
 * **`top_p`** (`float`, default 1.0): Nucleus sampling - limits to tokens with cumulative probability ≤ top\_p. Typical range: 0.1-1.0
 * **`top_k`** (`int`, default 40): Limits to top-k most probable tokens. Typical range: 1-100
 * **`repeat_penalty`** (`float`, default 1.0): Penalty for repeating tokens (>1.0 = discourage repetition). Typical range: 1.0-1.5
 
@@ -54,7 +54,7 @@
     messages=[{"role": "user", "content": "What is machine learning?"}],
     temperature=0.1,
     max_tokens=512,
     extra_body={"top_k": 50, "repeat_penalty": 1.05},
 )
 ```
 
@@ -134,7 +134,7 @@
 
 <Accordion title="Using the API">
   ```python
   from openai import OpenAI
   import base64
 
   client = OpenAI(
@@ -143,8 +143,8 @@
   )
 
   # Encode image to base64
   with open("image.jpg", "rb") as image_file:
       image_data = base64.b64encode(image_file.read()).decode("utf-8")
 
   response = client.chat.completions.create(
       model="<model-id-from-v1-models>",

deployment/on-device/llama-cpp.mdx

@@ -1,6 +1,6 @@
 ---
 title: "llama.cpp"
-description: "llama.cpp is a C++ library for efficient LLM inference with minimal dependencies. It's designed for CPU-first inference with cross-platform support."
+description: "Run Liquid AI LFM models locally with llama.cpp. Setup, GGUF downloads, and generation parameters for on-device inference."
 ---
 
 <Tip>
@@ -8,14 +8,14 @@
 </Tip>
 
 <Note>
   Building an application with llama.cpp? Jump to [Build with llama.cpp](#building-applications) for mobile, desktop, chat, tool-calling, structured-output, and multimodal guides.
 </Note>
 
 For GPU-accelerated inference at scale, consider using [vLLM](/deployment/gpu-inference/vllm) instead.
 
 <div className="colab-link">
   <a href="https://colab.research.google.com/github/Liquid4All/docs/blob/main/notebooks/LFM2_Inference_with_llama_cpp.ipynb" target="_blank">
     <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open in Colab" />
   </a>
 </div>
 
@@ -92,8 +92,8 @@
     ```bash
     git clone https://github.com/ggml-org/llama.cpp
     cd llama.cpp
     cmake -B build
     cmake --build build --config Release -j 8
     ```
 
     The compiled programs will be in `./build/bin/`.
@@ -215,7 +215,7 @@
 Control text generation behavior using parameters in the OpenAI-compatible API or command-line flags. Key parameters:
 
 * **`temperature`** (`float`): Controls randomness (0.0 = deterministic, higher = more random). Typical range: 0.1-2.0
 * **`top_p`** (`float`): Nucleus sampling - limits to tokens with cumulative probability ≤ top\_p. Typical range: 0.1-1.0
 * **`top_k`** (`int`): Limits to top-k most probable tokens. Typical range: 1-100
 * **`min_p`** (`float`): Filters tokens below `min_p * max_probability`. Typical range: 0.05-0.3
 * **`max_tokens`** / **`--n-predict`** (`int`): Maximum number of tokens to generate
@@ -247,7 +247,7 @@
 
 ## Vision Models
 
 LFM2-VL GGUF models can be used for multimodal inference with llama.cpp.
 
 ### Quick Start with llama-cli
 
@@ -307,7 +307,7 @@
 </Accordion>
 
 <Accordion title="Using llama-server">
   Start a vision model server with both the model and mmproj files:
 
   ```bash
   llama-server \
@@ -353,12 +353,12 @@
 </Accordion>
 
 <Info>
 For a complete working example with step-by-step instructions, see the [llama.cpp Vision Model Colab notebook](https://colab.research.google.com/drive/1q2PjE6O_AahakRlkTNJGYL32MsdUcj7b?usp=sharing).
 </Info>
 
 ## Converting Custom Models
 
 If you have a finetuned model or need to create a GGUF from a Hugging Face model:
 
 ```bash
 # Clone llama.cpp if you haven't already
@@ -377,7 +377,7 @@
 
 <CardGroup cols={2}>
   <Card title="iOS & Android" icon="mobile" href="/deployment/on-device/llama-cpp/mobile">
     Link the XCFramework or NDK build and run LFM GGUFs in-process.
   </Card>
   <Card title="Desktop & Server Apps" icon="desktop" href="/deployment/on-device/llama-cpp/desktop">
     llama-server as a sidecar, or Python / Node.js / .NET bindings.

deployment/on-device/lm-studio.mdx

@@ -1,6 +1,6 @@
 ---
 title: "LM Studio"
-description: "LM Studio is a desktop application for running LLMs locally with a graphical interface."
+description: "Run Liquid AI LFM models in LM Studio. Download GGUF models and start generating with a graphical interface."
 ---
 
 <Tip>
@@ -32,7 +32,7 @@
 Control text generation behavior using the GUI sidebar or API parameters. Key parameters:
 
 * **`temperature`** (`float`, default 1.0): Controls randomness (0.0 = deterministic, higher = more random). Typical range: 0.1-2.0
 * **`top_p`** (`float`, default 1.0): Nucleus sampling - limits to tokens with cumulative probability ≤ top\_p. Typical range: 0.1-1.0
 * **`top_k`** (`int`, default 40): Limits to top-k most probable tokens. Typical range: 1-100
 * **`max_tokens`** (`int`): Maximum number of tokens to generate
 * **`repeat_penalty`** (`float`, default 1.0): Penalty for repeating tokens (>1.0 = discourage repetition). Typical range: 1.0-1.5
@@ -46,7 +46,7 @@
     messages=[{"role": "user", "content": "What is machine learning?"}],
     temperature=0.1,
     max_tokens=512,
     extra_body={"top_k": 50, "repeat_penalty": 1.05},
 )
 ```
 
@@ -120,7 +120,7 @@
 
 <Accordion title="Using the API">
   ```python
   from openai import OpenAI
   import base64
 
   client = OpenAI(
@@ -129,8 +129,8 @@
   )
 
   # Encode image to base64
   with open("image.jpg", "rb") as image_file:
       image_data = base64.b64encode(image_file.read()).decode("utf-8")
 
   response = client.chat.completions.create(
       model="local-model",

deployment/on-device/mlx.mdx

@@ -1,13 +1,13 @@
 ---
 title: "MLX"
-description: "MLX is Apple's machine learning framework optimized for Apple Silicon. It provides efficient inference on Mac devices with M-series chips (M1, M2, M3, M4) using Metal acceleration for GPU computing."
+description: "Run Liquid AI LFM models on Apple Silicon with MLX. Setup, model downloads, and Metal-accelerated inference on M-series chips."
 ---
 
 <Tip>
   Use MLX for running models on Apple Silicon Macs with Metal GPU acceleration.
 </Tip>
 
 MLX leverages unified memory architecture on Apple Silicon, allowing seamless data sharing between CPU and GPU. The `mlx-lm` package provides a simple interface for loading and serving LLMs.
 
 ## Installation
 
@@ -47,7 +47,7 @@
 Control text generation behavior using parameters in the `generate()` function. Key parameters:
 
 * **`temperature`** (`float`, default 1.0): Controls randomness (0.0 = deterministic, higher = more random). Typical range: 0.1-2.0
 * **`top_p`** (`float`, default 1.0): Nucleus sampling - limits to tokens with cumulative probability ≤ top\_p. Typical range: 0.1-1.0
 * **`top_k`** (`int`, default 50): Limits to top-k most probable tokens. Typical range: 1-100
 * **`max_tokens`** (`int`): Maximum number of tokens to generate
 * **`repetition_penalty`** (`float`, default 1.0): Penalty for repeating tokens (>1.0 = discourage repetition). Typical range: 1.0-1.5
@@ -111,7 +111,7 @@
     ],
     temperature=0.3,
     max_tokens=512,
     extra_body={"min_p": 0.15, "repetition_penalty": 1.05},
 )
 print(response.choices[0].message.content)
 ```
@@ -134,11 +134,11 @@
 
 ## Vision Models
 
 LFM2-VL models support both text and image inputs for multimodal inference. Use `mlx_vlm` to load and generate with vision models:
 
 <Accordion title="Single Image Example">
   ```python
   from mlx_vlm import load, generate
   from mlx_vlm.prompt_utils import apply_chat_template
   from mlx_vlm.utils import load_image_processor
   from PIL import Image

deployment/on-device/ollama.mdx

@@ -1,21 +1,21 @@
 ---
 title: "Ollama"
-description: "Ollama is a command-line tool for running LLMs locally with a simple interface. It provides easy model management and serving with an OpenAI-compatible API."
+description: "Deploy Liquid AI LFM models with Ollama. Pull, configure, and serve LFM2.5 models locally with an OpenAI-compatible API."
 ---
 
 <Tip>
   Use Ollama for quick local model serving with a simple CLI or Docker-based deployment.
 </Tip>
 
 Ollama uses GGUF models and supports GPU acceleration (CUDA, Metal, ROCm).
 
 <Warning>
   The official Ollama v0.17.0 (latest stable) from [ollama.com](https://ollama.com) fails with a `missing tensor 'output_norm.weight'` error on the `lfm2moe` architecture. This affects all LFM MoE models (e.g. LFM2-24B-A2B, LFM2-8A-A1B). To run any LFM MoE model you specifically need [v0.17.1-rc0](https://github.com/ollama/ollama/releases/tag/v0.17.1-rc0) or later.
 </Warning>
 
 <div className="colab-link">
   <a href="https://colab.research.google.com/github/Liquid4All/docs/blob/main/notebooks/LFM2_Inference_with_Ollama.ipynb" target="_blank">
     <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open in Colab" />
   </a>
 </div>
 
@@ -52,15 +52,15 @@
     docker exec -it ollama ollama run hf.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF
     ```
 
     See the [Ollama Docker documentation](https://ollama.com/blog/ollama-is-now-available-as-an-official-docker-image) for more details.
   </Tab>
 </Tabs>
 
 ## Using LFM2 Models
 
 Ollama can load GGUF models directly from Hugging Face or from local files.
 
 ### Running GGUFs
 
 You can run LFM2 models directly from Hugging Face:
 
@@ -107,7 +107,7 @@
   PARAMETER stop "<|endoftext|>"
   ```
 
   Import the model with the Modelfile:
 
   ```bash
   ollama create my-model -f Modelfile
@@ -138,11 +138,11 @@
 ollama run hf.co/LiquidAI/LFM2.5-1.2B-Instruct-GGUF "What is machine learning?"
 ```
 
 If you imported a model with a custom name using a Modelfile, use that name instead (e.g., `ollama run my-model`).
 
 ## Serving Models
 
 Ollama automatically starts a server on `http://localhost:11434` with an OpenAI-compatible API for programmatic access.
 
 ### Python Client
 
@@ -166,7 +166,7 @@
 ```
 
 <Accordion title="Curl request examples">
   Ollama provides two native API endpoints:
 
   **Generate API** (simple completion):
 
@@ -191,7 +191,7 @@
 
 ## Vision Models
 
 LFM2-VL GGUF models can also be used for multimodal inference with Ollama.
 
 <Accordion title="Interactive Chat with Images">
   Run a vision model directly and provide images in the chat:
@@ -215,7 +215,7 @@
 
 <Accordion title="Using the API">
   ```python
   from openai import OpenAI
   import base64
 
   client = OpenAI(
@@ -224,8 +224,8 @@
   )
 
   # Encode image to base64
   with open("image.jpg", "rb") as image_file:
       image_data = base64.b64encode(image_file.read()).decode("utf-8")
 
   response = client.chat.completions.create(
       model="hf.co/LiquidAI/LFM2.5-VL-1.6B-GGUF",

deployment/on-device/onnx.mdx

@@ -1,13 +1,13 @@
 ---
 title: "ONNX"
-description: "ONNX provides a platform-agnostic inference specification that allows running the model on device-specific runtimes that include CPU, GPU, NPU, and WebGPU."
+description: "Run Liquid AI LFM models with ONNX Runtime on CPU, GPU, NPU, and WebGPU for cross-platform on-device inference."
 ---
 
 <Tip>
   Use ONNX for cross-platform deployment, edge devices, and browser-based inference with WebGPU and Transformers.js.
 </Tip>
 
 ONNX (Open Neural Network Exchange) is a portable format that enables LFM inference across diverse hardware and runtimes. ONNX models run on CPUs, GPUs, NPUs, and in browsers via WebGPU—making them ideal for edge deployment and web applications.
 
 ## LiquidONNX
 

deployment/tools/model-bundling/changelog.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Changelog"
+description: "Release history for leap-bundle, the Liquid AI model bundling CLI tool."
 ---
 
 ## `v0.9.0` - 2026-01-02

docs.json

@@ -32,6 +32,10 @@
       "vscode"
     ]
   },
+  "metadata": {
+    "og:site_name": "Liquid AI Docs",
+    "twitter:card": "summary"
+  },
   "favicon": "/favicon.svg",
   "logo": {
     "light": "/logo/light.svg",

examples/android/leap-koog-agent.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Build AI Agents with Koog Framework on Android"
+description: "Build on-device AI agents on Android using the Koog framework with Liquid AI LFM models, supporting tool calling and MCP."
 ---
 
 <Note>
@@ -10,7 +11,7 @@
   Browse the complete example on GitHub
 </Card>
 
 This example demonstrates how to integrate the **Koog framework** with LeapSDK on Android to build intelligent AI agents. Koog is an open-source framework for creating AI agents that can understand natural language, invoke tools, manage context, and integrate with MCP (Model Context Protocol) servers.
 
 LeapKoogAgent shows how to bring autonomous agent capabilities to mobile devices, enabling apps that can reason, plan, and execute complex tasks on behalf of users—all running locally on Android.
 
@@ -18,7 +19,7 @@
 
 LeapKoogAgent showcases advanced AI agent capabilities:
 
 - **Koog Framework Integration** - Build structured AI agents with reasoning capabilities
 - **Natural Language Understanding** - Process user intents and commands
 - **Tool Invocation** - Enable agents to call functions and use external tools
 - **Context Management** - Maintain conversation state and agent memory
@@ -29,9 +30,9 @@
 
 This example demonstrates building production-ready AI agents that run entirely on Android devices.
 
 ## What is the Koog Framework?
 
 **Koog** is an open-source framework for building AI agents. It provides a structured approach to creating agents that can:
 
 - **Reason and plan** - Break down complex tasks into steps
 - **Use tools** - Invoke functions to accomplish specific tasks (calculations, API calls, database queries)
@@ -57,7 +58,7 @@
 - Smart home controllers that understand natural language commands
 
 <Card title="Learn More About Koog" icon="book" href="https://docs.koog.ai">
   Explore the Koog framework documentation and examples
 </Card>
 
 ## Environment setup
@@ -101,7 +102,7 @@
   adb shell ls -lh /data/local/tmp/liquid/
   ```
 
   **Note:** Apps cannot read `/tmp/` on Android — use `/data/local/tmp/<your-namespace>/` for ADB-pushed assets (matches the other Android examples). If you deploy to a different location, update the `modelPath` in your app code accordingly. The example snippets below use `loadSimpleModel(model: ModelSource(...))` to load the sideloaded file; switch to `loadModel(modelName:, quantizationType:)` if you'd rather have the SDK download the model automatically.
 </Accordion>
 
 <Accordion title="Dependencies Setup">
@@ -153,7 +154,7 @@
    - Navigate to the `LeapKoogAgent` folder
 
 4. **Build the project**
    - Wait for Gradle sync to complete
    - Resolve any dependency issues
 
 5. **Run on device or emulator**
@@ -173,9 +174,9 @@
 
 ## Understanding the architecture
 
 ### Koog Agent Initialization
 
 Load the LEAP model first, then bridge it to a Koog agent. The Koog APIs below are illustrative — see the linked GitHub example for the canonical adapter wiring, since `ai.koog:koog-agents` evolves independently from LeapSDK.
 
 ```kotlin
 import ai.liquid.leap.ModelRunner
@@ -406,7 +407,7 @@
 
 ### Resource Cleanup
 
 **Important:** Always clean up agent resources properly to prevent memory leaks and ANRs:
 
 ```kotlin
 class AgentViewModel(application: Application) : AndroidViewModel(application) {
@@ -437,7 +438,7 @@
 ```
 
 **Why this matters:**
 - **Avoid `runBlocking`** in `onCleared()` - it blocks the main thread and can cause ANRs (Application Not Responding)
 - **Use `CoroutineScope(Dispatchers.IO).launch`** instead - model unloading happens asynchronously
 - **Always clean up agent resources** - prevents memory leaks and ensures proper shutdown
 - **Catch exceptions** - cleanup should never crash the app

examples/android/recipe-generator-constrained-output.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Generate Structured Recipes with Constrained Output"
+description: "Android app demonstrating constrained JSON generation to produce structured recipe output using a Liquid AI LFM model on-device."
 ---
 
 <Note>
@@ -10,7 +11,7 @@
   Browse the complete example on GitHub
 </Card>
 
 This example demonstrates **constrained generation** with LeapSDK on Android. Instead of generating free-form text, the RecipeGenerator app produces structured JSON output that follows a predefined schema, ensuring consistent and parseable results.
 
 Constrained generation is essential for building reliable AI-powered applications where the output needs to integrate with downstream systems or databases. This example shows how to enforce structure while maintaining creative and useful AI-generated content.
 
@@ -20,7 +21,7 @@
 
 - **Structured Output Generation** - Enforce JSON schema constraints on model outputs
 - **Automatic Model Downloading** - Models download automatically via `LeapDownloader` on first run
 - **Constrained Decoding** - Guide the model to produce valid, parseable data structures
 - **Schema Validation** - Ensure generated recipes follow a specific format
 - **Type Safety** - Parse AI outputs directly into Kotlin data classes with kotlinx.serialization
 - **@Generatable Annotation** - Use LeapSDK's annotation for simplified structured output generation
@@ -34,10 +35,10 @@
 **Constrained generation** refers to the ability to enforce specific formatting or structural requirements on model outputs. Instead of generating free-form text, the model produces outputs that conform to a predefined schema or pattern.
 
 **Benefits:**
 - **Reliability** - Guaranteed parseable output every time
 - **Type Safety** - Direct integration with typed programming languages
 - **Validation** - Automatic schema compliance
 - **Downstream Integration** - Feed structured data into databases, APIs, or UIs
 - **Error Reduction** - Eliminate parsing errors from inconsistent formats
 
 **Common use cases:**
@@ -150,8 +151,8 @@
    - Select "Open an existing project"
    - Navigate to the `RecipeGenerator` folder and open it
 
 3. **Gradle sync**
    - Wait for Gradle to sync all dependencies
    - Ensure LeapSDK 0.10.7 is downloaded
 
 4. **Run the app**
@@ -167,7 +168,7 @@
    - The model is cached for future use
 
 6. **Generate recipes**
    - Enter ingredients or a dish name (e.g., "pasta carbonara", "chocolate chip cookies")
    - Tap "Generate Recipe"
    - The app will produce a structured recipe with ingredients, steps, and metadata
    - All output will be valid JSON conforming to the recipe schema
@@ -176,7 +177,7 @@
   **After First Run**: The model is cached locally. Subsequent app launches work offline and start immediately without downloading.
 </Note>
 
 ## Code walkthrough
 
 The core business logic is implemented in `MainActivityViewModel.kt`. Here's how it works:
 
@@ -348,7 +349,7 @@
 ```
 
 **Best practices:**
 - Never use `runBlocking` in `onCleared()` - it causes ANRs (Application Not Responding)
 - Use `CoroutineScope(Dispatchers.IO).launch` for async cleanup
 - Always catch exceptions to prevent crashes during cleanup
 - This ensures smooth app shutdown without blocking the UI
@@ -422,7 +423,7 @@
     "In a separate bowl, whisk together flour, baking soda, and salt",
     "Gradually blend dry ingredients into butter mixture",
     "Stir in chocolate chips",
     "Drop rounded tablespoons of dough onto ungreased cookie sheets",
     "Bake for 9-11 minutes or until golden brown",
     "Cool on baking sheet for 2 minutes before transferring to a wire rack"
   ],

examples/android/slogan-generator.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Product Slogan Generator with LeapSDK"
+description: "Simple Android app for on-device text generation using a Liquid AI LFM model to create product slogans."
 ---
 
 <Note>
@@ -12,7 +13,7 @@
 
 This example demonstrates how to use LeapSDK for **single-turn generation** tasks on Android. The SloganApp generates creative product slogans on-device using local language models, showcasing how to integrate AI capabilities for specific product needs without requiring cloud connectivity.
 
 Built with traditional **Android Views**, this example provides a different UI approach compared to Jetpack Compose, making it ideal for teams working with legacy Android codebases.
 
 ## What's inside?
 
@@ -24,7 +25,7 @@
 - **Simple Integration** - Minimal setup for focused AI tasks
 - **Product-focused Prompting** - Optimized prompts for marketing and slogan generation
 
 This example is perfect for understanding how to integrate LeapSDK into existing Android applications that use the traditional View system rather than Jetpack Compose.
 
 ## Environment setup
 
@@ -77,8 +78,8 @@
    - Select "Open an existing project"
    - Navigate to the `SloganApp` folder and open it
 
 3. **Gradle sync**
    - Wait for Gradle to sync all dependencies
    - Resolve any dependency conflicts if prompted
 
 4. **Run the app**
@@ -126,7 +127,7 @@
 
 ### Traditional Android Views Approach
 
 Unlike examples that use Jetpack Compose, SloganApp demonstrates integration with the traditional Android View system:
 
 **XML Layout Structure:**
 ```xml

examples/android/vision-language-model-example.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Image Understanding with Vision Language Models"
+description: "Android app for on-device image understanding and visual Q&A using LFM2.5-VL-1.6B with Jetpack Compose."
 ---
 
 <Note>
@@ -10,18 +11,18 @@
   Browse the complete example on GitHub
 </Card>
 
 This example demonstrates how to use **Vision Language Models (VLMs)** with LeapSDK on Android. VLMs combine image understanding with natural language processing, enabling your app to analyze images, answer questions about visual content, and generate detailed descriptions—all on-device.
 
 Built with **Jetpack Compose** and the **Coil** image loading library, this example shows how to create a multimodal AI application that processes both images and text locally on Android devices.
 
 ## What's inside?
 
 The VLMExample showcases cutting-edge multimodal AI capabilities:
 
 - **Vision Language Models** - Analyze images and generate text descriptions
 - **Image Input Processing** - Handle image selection from gallery or camera
 - **Multimodal Understanding** - Combine visual and textual information
 - **Jetpack Compose UI** - Modern, declarative UI for image display and results
 - **Coil Integration** - Efficient image loading and rendering
 - **On-device Inference** - Complete privacy with local VLM processing
 - **Interactive Q&A** - Ask questions about images and get contextual answers
@@ -30,7 +31,7 @@
 
 ## What are Vision Language Models?
 
 **Vision Language Models (VLMs)** are AI models that can process both images and text simultaneously, enabling them to:
 
 - **Describe images** - Generate detailed captions of what's in a photo
 - **Answer visual questions** - Respond to queries about image content ("What color is the car?")
@@ -79,7 +80,7 @@
 
   The example uses **LFM2.5-VL-1.6B** from the [LEAP Model Library](https://leap.liquid.ai/models). The code below loads it via `LeapModelDownloader.loadModel(...)`, which fetches the GGUF + matching `mmproj` companion file from the registry on first launch and caches them on device. No manual ADB step is needed.
 
   **Manual deployment (alternative).** If you'd rather sideload the files, push the GGUF and its `mmproj` companion via ADB and use `loadSimpleModel(model: ModelSource(...))` instead of `loadModel(...)`:
 
   ```bash
   adb shell mkdir -p /data/local/tmp/liquid/
@@ -121,14 +122,14 @@
 
   [Coil](https://coil-kt.github.io/coil/) is a Kotlin-first image loading library for Android that:
   - Efficiently loads and caches images
   - Integrates seamlessly with Jetpack Compose
   - Handles image transformations and processing
   - Provides modern coroutine-based APIs
 </Accordion>
 
 ## How to run it
 
 Follow these steps to start analyzing images with VLMs:
 
 1. **Clone the repository**
    ```bash
@@ -138,7 +139,7 @@
 
 2. **Pick a deployment strategy**
    - Default: let `LeapModelDownloader.loadModel(...)` fetch LFM2.5-VL-1.6B from the LEAP Model Library on first launch
    - Optional manual push: see the **Model Bundle Setup** accordion above for the ADB sideload path
 
 3. **Open in Android Studio**
    - Launch Android Studio
@@ -286,7 +287,7 @@
 
 **Resource cleanup best practices:**
 - Always unload models in `onCleared()` to prevent memory leaks
 - Never use `runBlocking` in `onCleared()` - it causes ANRs
 - Use async cleanup with `CoroutineScope(Dispatchers.IO).launch`
 - Catch exceptions to ensure cleanup doesn't crash the app
 
@@ -399,13 +400,13 @@
 }
 ```
 
 `ModelRunner.unload()` is `suspend` (per `ai.liquid.leap.ModelRunner`), so call it from a coroutine scope.
 
 ## Results
 
 The VLMExample demonstrates powerful image understanding capabilities:
 
 ![VLMExample Screenshot](https://raw.githubusercontent.com/Liquid4All/LeapSDK-Examples/main/Android/VLMExample/docs/vlm_example.png)
 
 The interface shows:
 - Selected image displayed clearly with Coil

examples/android/web-content-summarizer.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Web Content Summarizer for Android"
+description: "Android app that summarizes web pages on-device using a Liquid AI LFM model, with no cloud dependencies."
 ---
 
 <Note>
@@ -126,8 +127,8 @@
    - Select "Open an existing project"
    - Navigate to the `ShareAI` folder and open it
 
 3. **Gradle sync**
    - Wait for Gradle to sync all dependencies
    - Ensure all libraries download successfully
 
 4. **Install on device**
@@ -186,7 +187,7 @@
 
 ### Web Content Extraction
 
 The app uses **Jsoup** to parse HTML and extract main content:
 
 ```kotlin
 suspend fun extractWebContent(url: String): String = withContext(Dispatchers.IO) {
@@ -265,11 +266,11 @@
 
 1. **Receive Intent** → Extract URL from share intent
 2. **Fetch Content** → Download HTML via OkHttp
 3. **Parse HTML** → Extract main text content with Jsoup
 4. **Generate Summary** → Process with LeapSDK on-device
 5. **Display Result** → Show summary in UI with options to copy/share
 
 All steps run asynchronously using Kotlin coroutines to keep the UI responsive.
 
 ## Demo
 
@@ -301,7 +302,7 @@
 **Privacy-conscious Reading**
 - Summarize sensitive content without sending it to cloud services
 - Process work-related articles without data leaving your device
 - Read summaries of paywalled content (from accessible portions)
 
 **Time Efficiency**
 - Decide if an article is worth reading in full
@@ -316,7 +317,7 @@
 - **Summary length control** - Let users choose brief, medium, or detailed summaries
 - **Language detection** - Automatically detect and handle non-English content
 - **Offline queue** - Save shared URLs for summarization when model is available
 - **Export options** - Save summaries as text files, PDFs, or to note-taking apps
 - **Reading time estimate** - Show how long the original article would take to read
 - **Key quotes extraction** - Highlight important quotes in addition to summary
 - **Topic tagging** - Automatically categorize summaries by topic

examples/customize-models/car-maker-identification.mdx

@@ -1,12 +1,13 @@
 ---
 title: "Fine tuning LFM2-VL to identify car makers from images"
+description: "Fine-tune LFM2-VL to identify car makers from images using structured output, LoRA, and an evaluation pipeline."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/car-maker-identification">
   Browse the complete example on GitHub
 </Card>
 
 A step by step guide on fine tuning a Vision Language Model for image identification tasks. The task we solve in this example is to identify the car maker from an image, but the learnings transfer to any other image classification task you might be interested in.
 
 ![](/images/examples/customize-models/assets/images/task-01484227e78d02da63eeb840de314523.png)
 
@@ -107,14 +108,14 @@
 ### Weights & Biases setup[​](#weights--biases-setup "Direct link to Weights & Biases setup")
 
 <Accordion title="Click to see installation instructions">
   1. Create an account at [wandb.ai](https://wandb.ai/)
   2. Install the Weights & Biases Python package inside your virtual environment:
      ```
      uv add wandb
      ```
   3. Authenticate with Weights & Biases:
      ```
      uv run wandb login
      ```
      This will open a browser window where you can copy your API key and paste it in the terminal.
 </Accordion>
@@ -124,7 +125,7 @@
 <Accordion title="Click to see installation instructions">
   1. Install make:
      ```
      sudo apt-get install make
      ```
 </Accordion>
 
@@ -146,7 +147,7 @@
 
 3. **Fine-tune with LoRA**. Apply parameter-efficient fine-tuning using Low-Rank Adaptation to improve model accuracy while keeping computational costs manageable.
 
 4. **Evaluate improvements**. Compare fine-tuned model performance against baselines to measure the effectiveness of our customization. If you are happy with th results, you are done. Otherwise, you need to dig deeper into model failures and improve the dataset you started with, or the fine-tuning process.
 
 Let's go one by one.
 
@@ -174,11 +175,11 @@
 * **Classes**: 49 unique car manufacturers.
 * **Splits**: Train (6,860 images) and test (6,750 images) splits.
 
 The dataset includes additional splits with various image corruptions (gaussian noise, motion blur, etc.) for robustness testing, making it ideal for evaluating model performance under different conditions. In this tutorial we will only use the train and test splits.
 
 ## Step 2. Baseline performance of LFM2-VL models[​](#step-2-baseline-performance-of-lfm2-vl-models "Direct link to Step 2. Baseline performance of LFM2-VL models")
 
 Before embarking into any fine-tuning experiment, we need to establish a baseline performance for existing models. In this case, we will evaluate the peformance of
 
 * LFM2-VL-450M
 * LFM2-VL-1.6B
@@ -195,7 +196,7 @@
 └── eval_lfm_3B_raw_generation.yaml
 ```
 
 These YAML configs are loaded into our Python script using the `EvaluationConfig` class in `src/car_maker_identification/config.py`, so we ensure
 
 * all necessary parameters are passed and
 * they all have valid values according to their types.
@@ -204,14 +205,14 @@
 
 <Accordion title="Click to see system and user prompts">
   ```yaml
   system_prompt: |
     You excel at identifying car makers from pictures.
 
   user_prompt: |
     What car maker do you see in this picture?
     Pick one from the following list:
     - AM
     - Acura
     - Aston
     - Audi
     - BMW
@@ -221,19 +222,19 @@
     - Cadillac
     - Chevrolet
     - Chrysler
     - Daewoo
     - Dodge
     - Eagle
     - FIAT
     - Ferrari
     - Fisker
     - Ford
     - GMC
     - Geo
     - HUMMER
     - Honda
     - Hyundai
     - Infiniti
     - Isuzu
     - Jaguar
     - Jeep
@@ -241,7 +242,7 @@
     - Land
     - Lincoln
     - MINI
     - Maybach
     - Mazda
     - McLaren
     - Mercedes-Benz
@@ -252,7 +253,7 @@
     - Ram
     - Rolls-Royce
     - Scion
     - Spyker
     - Suzuki
     - Tesla
     - Toyota
@@ -287,7 +288,7 @@
 
 However, before blaming it on the model, let's dig deeper into the model predictions to see what's going on.
 
 If you look at the confusion matrix of the 3B model, you will see that even though the model works okeyish overall, it sometimes fails terribly too, and generates output that does not correspond to any car maker name.
 
 ![](/images/examples/customize-models/assets/images/confusion_matrix_lfm2_3b_raw_generation-2b4949c6ff212ebbb275982abf4784fb.png)
 
@@ -362,7 +363,7 @@
 
 ![](/images/examples/customize-models/assets/images/confusion_matrix_lfm2_3b_structured_generation-f93ecf885ff6ee370009f1c97873de1b.png)
 
 At this point you need to decide if the performance is good enough for your use case. If it is, you are done. Otherwis, it is time to fine-tune the model.
 
 ## Step 4. Fine tuning with LoRA[​](#step-4-fine-tuning-with-lora "Direct link to Step 4. Fine tuning with LoRA")
 
@@ -387,17 +388,17 @@
 
 At this stage we know that for the same training dataset, the LFM2-VL-3B model is able to fit the data better than the other two models.
 
 Apart from this, the train loss curves do not tell us much about the actual model peformance. Language Models are highly parametric neural networks, that can fit **anything** in the training dataset. This **anything** includes both actual patterns in the data, and noise, that does not generalize to the test set.
 
 So, to get a better understanding of the model performance, we need to evaluate the model on a held-out dataset. This is what the evaluation loss curve tells us.
 
 ![](/images/examples/customize-models/assets/images/eval_loss-dc36b175d7b477b1a872968ee044c8f6.png)
 
 Again LFM2-VL-3B is able to fit the data better than the other two models. When you look at its evaluation loss curve, you can see that it is stricly decreasing, meaning the model is learning epoch by epoch and still has not gotten to the point of overfitting.
 
 The following table shows the evaluation loss at different checkpoints during fine-tuning:
 
 | Checkpoint | Train Loss | Eval Loss |
 | ---------- | ---------- | --------- |
 | 100        | 5.82       | 5.46      |
 | 200        | 0.16       | 0.20      |
@@ -406,7 +407,7 @@
 | 1000       | 0.008      | 0.005     |
 
 <Check>
   Overfitting is when a model learns the noise in the training data, and does not generalize to the test set. In other words, the training loss is decreasing, but the evaluation loss is increasing.
 </Check>
 
 At this point, we can conclude that LFM2-VL-3B is the most promising model for our use case. However, we still need to check its actual performance on the test set.

examples/customize-models/home-assistant.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Fine-tuning LFM for a local Home Assistant"
+description: "End-to-end guide to fine-tuning and deploying a local home assistant with Liquid AI LFM models, covering prototyping, synthetic data, and tool calling."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/home-assistant">
@@ -17,7 +18,7 @@
 1. Build a proof of concept for a fully local Home Assistant.
 2. Benchmark its tool-calling accuracy so you have a clear baseline to improve on.
 3. Generate synthetic data for model fine-tuning.
 4. Fine-tune the model on this synthetic data to maximise accuracy using serverless GPUs by Modal.
 5. Deploy the fine-tuned model in the app.
 
 ## Quickstart
@@ -59,7 +60,7 @@
 
 Let's go one by one.
 
 ### Install uv
 
 <Accordion title="Click to see installation instructions for your platform">
   **macOS/Linux:**
@@ -106,14 +107,14 @@
 ### Weights & Biases setup
 
 <Accordion title="Click to see installation instructions">
   1. Create an account at [wandb.ai](https://wandb.ai/)
   2. Install the Weights & Biases Python package:
      ```bash
      uv add wandb
      ```
   3. Authenticate:
      ```bash
      uv run wandb login
      ```
 </Accordion>
 
@@ -266,7 +267,7 @@
 ```
 
 1. **Blocklist in the prompt.** Every benchmark utterance is listed and the model is told not to reproduce them.
 2. **Filter by exact or substring match.** Catches verbatim copies.
 3. **Filter by trigram similarity.** Catches light paraphrases by measuring overlapping three-word chunks. Candidates with more than 50% overlap are discarded.
 4. **Agent cross-validation.** Each surviving candidate is run through the real agent. Only examples where the agent produces the expected tool call are kept.
 
@@ -287,7 +288,7 @@
 
 Fine-tuning adapts the base model to our specific task. We use LoRA (Low-Rank Adaptation), which injects a small set of trainable weight matrices on top of the frozen base model. This keeps GPU memory usage low and training fast while still producing meaningful accuracy gains.
 
 Training runs on [Modal](https://modal.com) (a serverless GPU cloud) via [leap-finetune](https://github.com/Liquid4All/leap-finetune), Liquid AI's open source fine-tuning tool.
 
 ```mermaid
 sequenceDiagram
@@ -358,7 +359,7 @@
 | lights | 25.0% | 87.5% |
 | scene | 0.0% | 80.0% |
 | doors | 56.2% | 56.2% |
 | multi_tool | 8.3% | 33.3% |
 | status | 0.0% | 30.0% |
 | thermostat | 0.0% | 18.8% |
 | rejection | 0.0% | 0.0% |

examples/customize-models/satellite-vlm.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Fine-Tune a Vision-Language Model on Satellite Imagery"
+description: "Fine-tune LFM2.5-VL-450M on satellite imagery for visual question answering, grounding, and captioning with leap-finetune."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/satellite-vlm">
@@ -9,7 +10,7 @@
 Fine-tune [LiquidAI/LFM2.5-VL-450M](https://huggingface.co/LiquidAI/LFM2.5-VL-450M) on satellite imagery tasks using
 
 - [leap-finetune](https://github.com/Liquid4All/leap-finetune/): Liquid AI's fine-tuning framework for LFM models.
 - [Modal](https://modal.com): serverless GPU cloud for running data prep and training without managing infrastructure.
 - [VRSBench](https://huggingface.co/datasets/xiang709/VRSBench) dataset (NeurIPS 2024) which supports three tasks:
   - **VQA**: answer questions about satellite images (123K QA pairs)
   - **Visual Grounding**: detect and localize objects with bounding boxes (52K references)
@@ -21,7 +22,7 @@
 
 1. **Install and authenticate** (run from `cookbook/examples/satellite-vlm/`):
 
    Modal provides serverless GPUs you pay for per second. New accounts include $30 of free credit, enough to run this example end to end. No local GPU required.
 
    ```bash
    uv sync
@@ -48,13 +49,13 @@
 
 All heavy computation runs in the cloud. The only things that run locally are the `prepare_vrsbench.py` launcher and the `leap-finetune` CLI, both of which just submit jobs and stream logs.
 
 1. **Data prep (Modal CPU container):** `prepare_vrsbench.py --modal` spins up a Modal container, downloads VRSBench (~12 GB) from HuggingFace, converts it to JSONL, and writes everything to a Modal volume named `satellite-vlm`.
 2. **Training (Modal H100):** `leap-finetune` submits a training job that reads data from the same volume, fine-tunes the model, and saves checkpoints back to the volume.
 3. **Retrieval (local):** you pull the checkpoints from the volume to your local machine with `modal volume get`.
 
 ## Data Preparation
 
 `prepare_vrsbench.py` downloads VRSBench from HuggingFace and converts it to the JSONL format required by leap-finetune.
 
 **Modal (recommended):** runs entirely in the cloud, writes directly to the `satellite-vlm` Modal volume:
 
@@ -82,7 +83,7 @@
 
 The job runs on an H100, streams logs to your terminal, and saves checkpoints to the `satellite-vlm` Modal volume under `/satellite-vlm/outputs/`.
 
 To enable experiment tracking, uncomment `tracker: "wandb"` in the config.
 
 ## Retrieving Checkpoints
 
@@ -95,7 +96,7 @@
 
 ## Data Format
 
 The grounding task uses JSON bounding box format with 0-1 normalized coordinates, matching the LFM VLM's pretraining format:
 
 ```
 User:      Inspect the image and detect the large white ship.
@@ -112,21 +113,21 @@
 
 Benchmarks run automatically during training at every `eval_steps`:
 
 - **VQA**: `short_answer` metric (case-insensitive substring match)
 - **Grounding**: `grounding_iou` metric (IoU@0.5 threshold)
 - **Captioning**: `CIDEr` or `BLEU` metrics
 
 Each eval dataset can be limited (e.g., 500 samples) via the `limit` field in the YAML config for faster iteration.
 
 **Running a full standalone evaluation:** to evaluate on the complete dataset without retraining, use `configs/vrsbench_full_eval.yaml`. Set `eval_on_start: true`, remove the `limit` fields, and point the config at your checkpoint path. The model runs the full evaluation at step 0, logs results to WandB, and terminates.
 
 ## AI in Space Hackathon
 
 This example is the official starting point for the [**AI in Space Hackathon**](https://luma.com/n9cw58h0), a fully online event organised in partnership between [DPhi Space](https://github.com/DPhi-Space/SimSat) and Liquid AI, open to builders from all around the globe.
 
 Use this fine-tuning pipeline as your baseline and push it further with real satellite data.
 
 [![AI in Space Hackathon](/images/examples/customize-models/satellite-vlm/hackathon_banner.jpeg)](https://luma.com/n9cw58h0)
 
 <CardGroup cols={1}>
 

examples/customize-models/wildfire-prevention.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Build a Wildfire Prevention System with a Compact VLM"
+description: "Build an on-satellite wildfire detection system using Sentinel-2 imagery and LFM2.5-VL-450M for in-orbit inference."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/wildfire-prevention">
@@ -9,7 +10,7 @@
 In this example you will learn how to build a basic wildfire prevention system using:
 
 - Sentinel-2 satellite images
 - [LFM2.5-VL-450M](https://huggingface.co/LiquidAI/LFM2.5-VL-450M), a compact Vision-Language Model running directly on the satellite, so inference happens in orbit and only a lightweight JSON payload is downlinked to Earth.
 
 ```mermaid
 flowchart LR
@@ -50,9 +51,9 @@
 <Note>
   **What is Sentinel-2?**
 
   Sentinel-2 is a European Space Agency (ESA) satellite mission that captures high-resolution optical imagery of Earth's surface. It's part of the EU's Copernicus programme.
 
   It consists of 3 satellites (Sentinel-2A, 2B and 2C) which orbit in tandem, revisiting the same location every 5 days at the equator (more frequently at higher latitudes), and capturing **multispectral** images across **13 discrete wavelength ranges** simultaneously. Each range is called a band, and each band carries information about vegetation health, water content, soil moisture, or atmospheric conditions that is not visible to the naked eye.
 </Note>
 
 In this repository we use two different images for a given location:
@@ -76,15 +77,15 @@
    ```json
    {
      "risk_level": "high",
      "dry_vegetation_present": true,
      "urban_interface": true,
      "steep_terrain": true,
      "water_body_present": false,
      "image_quality_limited": false
    }
    ```
 
 3. This payload is downlinked to ground control on Earth. As the image tile has high risk, the system sends an alert to local fire services. These can then take precautionary measures like ground patrol deployment or controlled burns to reduce available fuel.
 
 ## 2. System design
 
@@ -94,7 +95,7 @@
 
 The bottleneck is not capability. It is **data transmission**.
 
 A frontier model runs on a server on Earth. To use it, the satellite downlinks raw images to a ground station, the ground station feeds the model, and the model produces the output on Earth. Images are high-dimensional: large matrices of pixel values per band, per frame. Multiply that by the number of captures per orbit, and you have a serious bandwidth problem.
 
 ```mermaid
 flowchart LR
@@ -112,7 +113,7 @@
     A =="|⚠️ raw image downlink<br/>hundreds of MB per frame|"==> B
 ```
 
 A small model removes that bottleneck entirely. At 450M parameters, LFM2.5-VL-450M is compact enough to run directly on the satellite: the satellite captures the image and runs inference on-board, the local model produces the payload output in orbit, and only the lightweight output is downlinked to the ground station.
 
 ```mermaid
 flowchart LR
@@ -138,26 +139,26 @@
   |----|----------|
   | `angeles_nf_ca` | Angeles National Forest, California |
   | `santa_barbara_ca` | Santa Barbara, California |
   | `napa_valley_ca` | Napa Valley, California |
   | `sierra_nevada_ca` | Sierra Nevada, California |
   | `alentejo_portugal` | Alentejo, Portugal |
   | `attica_greece` | Attica, Greece |
   | `cerrado_brazil` | Cerrado, Brazil |
   | `patagonia_argentina` | Patagonia, Argentina |
   | `black_forest_germany` | Black Forest, Germany |
   | `scottish_highlands` | Scottish Highlands |
   | `borneo_rainforest` | Borneo Rainforest |
   | `tanzania_savanna` | Tanzania Savanna |
   | `outback_nsw_australia` | Outback NSW, Australia |
   | `victorian_alpine_au` | Victorian Alps, Australia |
   | `kalahari_botswana` | Kalahari, Botswana |
   | `zagros_iran` | Zagros Mountains, Iran |
   | `negev_israel` | Negev Desert, Israel |
   | `alpine_switzerland` | Swiss Alps |
   | `amazon_brazil` | Amazon, Brazil |
   | `congo_basin_drc` | Congo Basin, DRC |
   | `lahaina_maui_hi` | Lahaina, Maui, Hawaii |
   | `mati_attica_gr` | Mati, Attica, Greece |
 </Accordion>
 
 ### Quickstart
@@ -245,7 +246,7 @@
 2. Samples `--n-temporal-tiles` timestamps evenly spaced within `[--start-date, --end-date]` using bin-center placement, so timestamps are always in the interior of the window.
 3. Builds a centered square grid of `--n-spatial-tiles` tiles around each location center, spaced `--size-km` apart.
 4. Fetches the RGB and SWIR images in parallel from SimSat for each `(spatial tile, timestamp)` pair.
 5. Saves `rgb.png` and `swir.png` to the tile subfolder.
 6. Sends both images to `claude-opus-4-6` for risk annotation, with automatic retry on rate-limit errors.
 7. Saves the structured JSON output as `annotation.json`.
 8. Assigns the tile to `train/` or `test/` based on a temporal cutoff. This prevents near-duplicate images (Sentinel-2 revisits every 5 days) from appearing on both sides of the split.
@@ -314,20 +315,20 @@
 - `results.json`: per-sample records with the model's actual predictions, ground truth, and per-field match results
 - `meta.json`: run metadata (model, dataset, backend, split)
 
 Once you have two or more eval runs, launch the comparison app to explore results visually:
 
 ```bash
 uv run streamlit run app/eval_compare.py
 ```
 
 ![Eval comparison app demo](/images/examples/customize-models/wildfire-prevention/eval_app_demo.gif)
 
 ### Results
 
 Evaluated on 22 locations ([Paulescu/wildfire-prevention](https://huggingface.co/datasets/Paulescu/wildfire-prevention)), ground truth from `claude-opus-4-6`:
 
 - `claude-opus-4-6` scores 0.99 overall, near-perfect across all fields. The result is not 100% due to non-determinism in token sampling.
 - The base LFM2.5-VL-450M scores 0.38 overall: it produces valid JSON reliably but struggles with field accuracy, especially `risk_level` (0.08) and `urban_interface` (0.25). This is expected for a zero-shot compact model on a specialized task. Fine-tuning addresses this gap.
 
 | field | claude-opus-4-6 | LFM2.5-VL-450M Q8\_0 |
 |---|---|---|
@@ -344,7 +345,7 @@
 
 ## 5. Fine-tuning
 
 We use [leap-finetune](https://github.com/Liquid4All/leap-finetune) to fine-tune `LFM2.5-VL-450M` on the Opus-labeled dataset via Modal's serverless H100 infrastructure.
 
 ### Step 1. Install leap-finetune
 
@@ -421,8 +422,8 @@
 
 Two important observations:
 
 - **Full fine-tuning, not LoRA (`use_peft: false`):** we update both the multimodal projector and the full language model backbone. Satellite imagery is severely underrepresented in standard VLM pretraining data, so the projector needs to genuinely re-learn how to map multispectral patches into meaningful tokens. At 450M parameters, full fine-tuning fits on a single H100 without the memory pressure that motivates LoRA on larger models.
 - **Modal section:** the `modal` block tells leap-finetune to run the training job on Modal's serverless GPU platform rather than locally. It specifies the GPU type (`H100:1`), a timeout, and the Modal volume where the prepared dataset lives and where checkpoints are written.
 
 ### Step 4. Kick off the fine-tuning
 
@@ -470,9 +471,9 @@
 ```
 
 - **`--output`** sets the backbone path: the language model weights, quantized to Q8\_0 by default.
 - The mmproj (`mmproj-lfm2.5-vl-wildfire-Q8_0.gguf`) is written automatically to the same directory, with `mmproj-` prepended. It contains the vision tower and multimodal projector weights (always F16).
 
 To use a different quantization level for the backbone, pass `--quant Q4_K_M` (or `Q4_0`, `Q5_K_M`, `Q6_K`, `F16`). The mmproj is always F16 regardless of `--quant`.
 
 ### Step 7. (Optional) Push the GGUF pair to HuggingFace
 
@@ -505,7 +506,7 @@
 
 ### Results
 
 Evaluated on 172 test samples ([Paulescu/wildfire-prevention](https://huggingface.co/datasets/Paulescu/wildfire-prevention)), ground truth from `claude-opus-4-6`. Fine-tuning takes the model from 0.38 to 0.84 overall accuracy, more than doubling performance. The largest gains are on `risk_level` (0.08 → 0.76), `urban_interface` (0.25 → 0.93), and `image_quality_limited` (0.28 → 0.86).
 
 | field | claude-opus-4-6 | LFM2.5-VL-450M Q8\_0 (base) | LFM2.5-VL-450M Q8\_0 (fine-tuned) |
 |---|---|---|---|

examples/index.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Examples Library"
+description: "Browse example projects and tutorials built with Liquid AI LFM models across laptop, Android, web, and fine-tuning use cases."
 ---
 
 ## Laptop
@@ -61,11 +62,11 @@
 </Card>
 
 <Card title="Vision Language Model Demo" icon="eye" href="/examples/android/vision-language-model-example">
   Analyze images and answer visual questions on Android using Vision Language Models. Built with Jetpack Compose and Coil.
 </Card>
 
 <Card title="AI Agents with Koog" icon="robot" href="/examples/android/leap-koog-agent">
   Build intelligent AI agents on Android with the Koog framework. Demonstrates tool invocation, context management, and MCP integration.
 </Card>
 
 </CardGroup>

examples/laptop-examples/audio-car-cockpit.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Audio car cockpit demo"
+description: "Voice-controlled car cockpit demo combining LFM2.5-Audio for speech and LFM2 for tool calling, running entirely local."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/audio-car-cockpit">
@@ -8,7 +9,7 @@
 
 ## What's inside?
 
 Combines [LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B-GGUF) in TTS and STT modes with [LFM2-1.2B-Tool](https://huggingface.co/LiquidAI/LFM2-1.2B-Tool-GGUF) within a mockup of a car cockpit, letting the user control the car functionalities by voice.
 
 All running locally in real-time.
 

examples/laptop-examples/audio-to-text-in-real-time.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Audio transcription in real-time"
+description: "Real-time audio transcription using LFM2.5-Audio-1.5B with llama.cpp, running fully local on your machine."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/audio-transcription-cli">
@@ -18,7 +19,7 @@
     cd cookbook/examples/audio-transcription-cli
     ```
 
 2. Install uv on your system, if you don't have it already.
 
     <Accordion title="Click to see installation instructions for uv">
       **macOS/Linux:**
@@ -79,7 +80,7 @@
 
 ## llama.cpp support for audio models
 
 [llama.cpp](https://github.com/ggerganov/llama.cpp) is a super fast and lightweight open-source inference engine for Language Models. It is written in C++ and can be used to run LLMs on your local machine. For example, our Python CLI used llama.cpp under the hood to deliver fast transcriptions, instead of using either `PyTorch` or the higher-level `transformers` library.
 
 In the [examples.sh](https://github.com/Liquid4All/cookbook/blob/main/examples/audio-transcription-cli/examples.sh) script you will find 3 examples on how to run inference with LFM2.5-Audio-1.5B for 3 common use cases:
 

examples/laptop-examples/browser-control.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Browser control with GRPO reinforcement learning"
+description: "Train a Liquid AI LFM model for web automation using GRPO reinforcement learning with BrowserGym environments."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/browser-control">
@@ -16,7 +17,7 @@
 
 - A Vision Language Model (VLM) can take as inputs a screenshot and the user goal, and generates a sequence of actions to accomplish that goal.
 
     ![Browser control multimodal](https://github.com/Liquid4All/cookbook/raw/main/examples/browser-control/media/browser_control_multimodal.gif)
 
 - A text-only Language Model can take as input the HTML code of the page and proceeds in the same way.
 
@@ -46,7 +47,7 @@
 
 ### Example
 
 Task: "book the shortest one-way flight from LAF to Akuton, AK on 12/06/2016"
 
 ![Book flight](https://github.com/Liquid4All/cookbook/raw/main/examples/browser-control/media/book-flight.png)
 
@@ -101,35 +102,35 @@
 
 It is simple enough to showcase the whole training pipeline without having to spend too much time training the model.
 
 Feel free to choose another task from [this list](https://miniwob.farama.org/environments/list/) and burn some GPUs to hit good performance.
 
 ### 2. The RL algorithm -> GRPO
 
 As the Language Model interacts with the environment you collect a sequence of rollouts with sparse rewards from the environment. Using these sparse rewards the RL algorithm (e.g. GRPO) adjusts the Language Model parameters to increase the chances of getting larger rewards.
 
 Here is a sample of 4 rollouts for the same initial prompt/task, where the model:
 
 1. Solves the task in the first step
 2. Solves the task in the second step
 3. Solves the task in the third step
 4. Does not solve the task after 4 steps, which is the max rollout length we set.
 
 ![4 rollouts](https://github.com/Liquid4All/cookbook/raw/main/examples/browser-control/media/4_rollouts.gif)
 
 GRPO uses the relative performance within that group to determine which actions to reinforce.
 
 ![GRPO rewards](https://github.com/Liquid4All/cookbook/raw/main/examples/browser-control/media/grpo_rewards.jpg)
 
 Responses that perform better than their groupmates get positive advantages, while worse ones get negative advantages. This approach is more memory-efficient than previous RL algorithms like Proximal Policy Optimization (PPO) since it eliminates the need to train and store a second language model for the value function.
 
 GRPO has become one of the most popular algorithms used by AI labs and AI practitioners to train/fine-tune LMs with Reinforcement Learning.
 
 ### 3. The policy -> LFM2-350M
 
 We will be using [LFM2-350M](https://huggingface.co/LiquidAI/LFM2-350M), which is a small-yet-powerful-and-fast language model with:
 
 - knowledge (MMLU, GPQA)
 - instruction following (IFEval, IFBench)
 - mathematical reasoning (GSM8K, MGSM)
 
 To speed up training we will also add support for LoRA adapters, so we don't need a whole model fine-tune.
@@ -138,8 +139,8 @@
 
 The 3 components of our RL training framework are:
 
 1. The `GRPOTrainer` from the trl library that implements the GRPO algorithm. It needs to run on GPU so the backward pass that updates model parameters is fast enough.
 2. The `vLLM` server to generate rollouts with LFM2-350M. It needs to run on GPU to parallelize the generation of rollouts.
 3. The Docker container with the `BrowserGym` environment, running as a Hugging Face space. Feel free to run on locally or on your infra Kubernetes. This one can run on CPU.
 
 ![3 components](https://github.com/Liquid4All/cookbook/raw/main/examples/browser-control/media/3_components.jpg)
@@ -148,7 +149,7 @@
 
 ![Collocate vLLM](https://github.com/Liquid4All/cookbook/raw/main/examples/browser-control/media/collocate_vllm.jpg)
 
 We use [Modal](https://modal.com/) for serverless GPU infrastructure. Modal is a great option for AI engineers who don't want to worry too much about infrastructure and prefer to pay per usage. It provides on-demand GPU access with simple Python-based configuration, making it easy to scale training workloads without managing clusters or dealing with complex DevOps setup.
 
 ## Experiments
 

examples/laptop-examples/build-a-game-with-decision-models.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Build a game with decision models"
+description: "Build a pixel-art survival racer where Liquid AI decision models choose the safest lane in real time via structured classification."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/road-decider">
@@ -10,7 +11,7 @@
 
 ![Road Decider demo](/images/examples/laptop-examples/road-decider/demo.png)
 
 Two cars race side by side on identical roads that get faster over time. Each car has three lives, and hitting traffic costs one. Play **You vs d1** with the arrow keys, or watch **Jev vs d1** to compare two decision models head-to-head.
 
 Decision models are purpose-built for classification, routing, and scoring. Instead of generating text token by token, they return a structured answer in a single call. That makes them fast enough to use inside a game loop, and cheap enough to call on every tick.
 
@@ -44,7 +45,7 @@
 ```
 
 <Note>
   Jev vs d1 mode is optional. To enable it, add an OpenRouter key as `OPENROUTER_API_KEY`.
 </Note>
 
 ### 3. Install dependencies and start the demo

examples/laptop-examples/flight-search-assistant.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Flight search assistant with tool calling"
+description: "CLI flight search assistant using LFM2.5-1.2B-Thinking for multi-step reasoning and tool calling to search and book flights."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/flight-search-assistant">
@@ -12,7 +13,7 @@
 
 ## Quickstart
 
 ### 1. Verify uv installation
 
 Check that you have `uv` available on your system:
 
@@ -57,7 +58,7 @@
 ```
 
 <Note>
   The implementation uses mocked flight data for demonstration purposes, though it supports integration with real APIs like Amadeus, Skyscanner, or Kiwi for production deployment.
 </Note>
 
 ## Future Development

examples/laptop-examples/invoice-extractor-tool-with-liquid-nanos.mdx

@@ -1,12 +1,13 @@
 ---
 title: "Invoice extractor tool"
+description: "Extract structured payment details from invoice PDFs using LFM2.5-VL-1.6B and llama.cpp, running fully local."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/invoice-parser">
   Browse the complete example on GitHub
 </Card>
 
 A Python CLI that extracts payment details from invoice PDFs.
 
 ![](https://github.com/Liquid4All/cookbook/raw/main/examples/invoice-parser/media/diagram.jpg)
 

examples/laptop-examples/lfm2-english-to-korean.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Bidirectional English to Korean translation CLI"
+description: "CLI translation tool using a fine-tuned Liquid AI LFM2 model for bidirectional English-Korean translation with automatic language detection."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/lfm2-english-to-korean">
@@ -8,7 +9,7 @@
 
 ## What's inside?
 
 An efficient bidirectional translation system powered by Liquid AI's LFM2 1.2B model fine-tuned for Korean-English translation. This project demonstrates how domain-specific fine-tuning can achieve superior performance compared to models 3x larger, outperforming Google's Gemma-3 4B and Alibaba's Qwen3 4B on the Flores-200 benchmark.
 
 Key features:
 - **Automatic language detection** - Intelligently detects input language and translates accordingly
@@ -48,7 +49,7 @@
 
 - **Base Model**: `gyung/lfm2-1.2b-koen-mt-v4-100k` - SFT fine-tuned LFM2 1.2B
 - **Adapter**: `gyung/lfm2-1.2b-koen-mt-v5-rl-10k-adapter` - LoRA adapter trained with GRPO
 - **Automatic Detection**: Regular expression pattern matching for Korean text (Hangul syllables, Jamo)
 
 The system automatically detects the input language and applies the appropriate translation direction, supporting both English→Korean and Korean→English translation.
 
@@ -103,7 +104,7 @@
 - **Speed optimization** with quantization techniques (GGUF, AWQ, GPTQ)
 - **llama.cpp integration** for faster CPU inference
 - **Full parameter RL training** with expanded compute resources
 - **Length normalization removal** based on recent Qwen team findings
 - **Extended dataset training** with 200K SFT + 25K RL samples
 
 ### Performance Optimization

examples/laptop-examples/meeting-summarization.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Meeting summarization CLI"
+description: "Summarize meeting transcripts locally using LFM2-2.6B-Transcript and llama.cpp with no cloud dependencies."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/meeting-summarization">
@@ -22,7 +23,7 @@
 
 ## Quick Start
 
 ### Install uv
 
 <Accordion title="Click to see installation instructions for uv">
   **macOS/Linux:**

examples/web/audio-webgpu-demo.mdx

@@ -1,5 +1,6 @@
 ---
 title: "LFM2.5-Audio browser demo with WebGPU"
+description: "Run LFM2.5-Audio-1.5B in the browser with WebGPU for speech recognition, text-to-speech, and audio conversations, fully client-side."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/audio-webgpu-demo">
@@ -61,12 +62,12 @@
 **2. Text-to-Speech (TTS)**
 - Input: Written text
 - Output: Natural-sounding audio
 - Use case: Create voice assistants, audiobooks, or accessibility features
 
 **3. Interleaved Mode**
 - Input: Mixed audio and text
 - Output: Conversational responses in text or audio
 - Use case: Interactive voice assistants and chatbots
 
 ## System Requirements
 

examples/web/hand-voice-racer.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Hand & Voice Racer"
+description: "Browser-based driving game controlled by hand gestures and voice commands using LFM2.5-Audio, running entirely client-side with WebGPU."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/hand-voice-racer">

examples/web/vl-webgpu-demo.mdx

@@ -1,5 +1,6 @@
 ---
 title: "Real-time video captioning with LFM2.5-VL-1.6B and WebGPU"
+description: "Run LFM2.5-VL-1.6B in the browser with WebGPU for real-time, private video captioning with no server required."
 ---
 
 <Card title="View Source Code" icon="github" href="https://github.com/Liquid4All/cookbook/tree/main/examples/vl-webgpu-demo">
@@ -77,9 +78,9 @@
 This demo can be deployed to any platform that supports CORS headers and SharedArrayBuffer:
 
 - **Hugging Face Spaces** - Recommended for quick deployment
 - **PaaS Providers** - Vercel, Netlify
 - **Cloud Storage** - AWS S3 + CloudFront, Google Cloud Storage, Azure Blob Storage
 - **Traditional Servers** - nginx, Apache, Caddy
 
 <Note>
   **Important**: GitHub Pages is not supported due to CORS and SharedArrayBuffer requirements.