VLM As Classifier¶

v2¶

Parse JSON strings from Visual Language Models (VLMs) and Large Language Models (LLMs) into standardized classification prediction format by extracting class predictions, mapping class names to class IDs, handling both single-class and multi-label formats, and converting VLM/LLM text outputs into workflow-compatible classification results for VLM-based classification, LLM classification parsing, and text-to-classification conversion workflows.

How This Block Works¶

This block converts VLM/LLM text outputs containing classification predictions into standardized classification prediction format. The block:

1. Receives image and VLM output string containing classification results in JSON format
2. Parses JSON content from VLM output:

Handles Markdown-wrapped JSON: - Searches for JSON wrapped in Markdown code blocks (json ...) - This format is common in LLM/VLM responses - If multiple markdown JSON blocks are found, only the first block is parsed - Extracts JSON content from within markdown tags

Handles raw JSON strings: - If no markdown blocks are found, attempts to parse the entire string as JSON - Supports standard JSON format strings 3. Detects classification format and parses accordingly:

Single-Class Classification Format: - Detects format containing "class_name" and "confidence" fields - Extracts the predicted class name and confidence score - Creates classification prediction with single top class - Maps class name to class ID using provided classes list

Multi-Label Classification Format: - Detects format containing "predicted_classes" array - Extracts all predicted classes with their confidence scores - Handles duplicate classes by taking maximum confidence - Maps all class names to class IDs using provided classes list 4. Creates class name to class ID mapping: - Uses the provided classes list to create index mapping (class_name → class_id) - Maps classes in order (first class = ID 0, second = ID 1, etc.) - Classes not in the provided list get class_id = -1 5. Normalizes confidence scores: - Scales confidence values to valid range [0.0, 1.0] - Clamps values outside the range to 0.0 or 1.0 6. Constructs classification prediction: - Includes image dimensions (width, height) from input image - For single-class: includes "top" class, confidence, and predictions array - For multi-label: includes "predicted_classes" list and predictions dictionary - Includes inference_id and parent_id for tracking - Formats prediction in standard classification prediction format 7. Handles errors: - Sets error_status to True if JSON parsing fails - Sets error_status to True if classification format cannot be determined - Returns None for predictions when errors occur - Always includes inference_id for tracking 8. Returns classification prediction: - Outputs predictions in standard classification format (compatible with classification blocks) - Outputs error_status indicating parsing success/failure - Outputs inference_id with specific type for tracking and lineage

The block enables using VLMs/LLMs for classification by converting their text-based JSON outputs into standardized classification predictions that can be used in workflows like any other classification model output.

Common Use Cases¶

• VLM-Based Classification: Use Visual Language Models for image classification by parsing VLM outputs into classification predictions (e.g., classify images with VLMs, use GPT-4V for classification, parse Claude Vision classifications), enabling VLM classification workflows
• LLM Classification Parsing: Parse LLM text outputs containing classification results into standardized format (e.g., parse GPT classification outputs, convert LLM predictions to classification format, use LLMs for classification), enabling LLM classification workflows
• Text-to-Classification Conversion: Convert text-based classification outputs from models into workflow-compatible classification predictions (e.g., convert text predictions to classification format, parse text-based classifications, convert model outputs to classifications), enabling text-to-classification workflows
• Multi-Format Classification Support: Handle both single-class and multi-label classification formats from VLM/LLM outputs (e.g., support single-label VLM classifications, support multi-label VLM classifications, handle different classification formats), enabling flexible classification workflows
• VLM Integration: Integrate VLM outputs into classification workflows (e.g., use VLMs in classification pipelines, integrate VLM predictions with classification blocks, combine VLM and traditional classification), enabling VLM integration workflows
• Flexible Classification Sources: Enable classification from various model types that output text/JSON (e.g., use any text-output model for classification, convert model outputs to classifications, parse various classification formats), enabling flexible classification workflows

Connecting to Other Blocks¶

This block receives images and VLM outputs and produces classification predictions:

• After VLM/LLM blocks to parse classification outputs into standard format (e.g., VLM output to classification, LLM output to classification, parse model outputs), enabling VLM-to-classification workflows
• Before classification-based blocks to use parsed classifications (e.g., use parsed classifications in workflows, provide classifications to downstream blocks, use VLM classifications with classification blocks), enabling classification-to-workflow workflows
• Before filtering blocks to filter based on VLM classifications (e.g., filter by VLM classification results, use parsed classifications for filtering, apply filters to VLM predictions), enabling classification-to-filter workflows
• Before analytics blocks to analyze VLM classification results (e.g., analyze VLM classifications, perform analytics on parsed classifications, track VLM classification metrics), enabling classification analytics workflows
• Before visualization blocks to display VLM classification results (e.g., visualize VLM classifications, display parsed classification predictions, show VLM classification outputs), enabling classification visualization workflows
• In workflow outputs to provide VLM classifications as final output (e.g., VLM classification outputs, parsed classification results, VLM-based classification outputs), enabling classification output workflows

Version Differences¶

This version (v2) includes the following enhancements over v1:

• Improved Type System: The inference_id output now uses INFERENCE_ID_KIND instead of generic STRING_KIND, providing better type safety and semantic clarity for inference ID values in the workflow type system

Requirements¶

This block requires an image input (for metadata and dimensions) and a VLM output string containing JSON classification data. The JSON can be raw JSON or wrapped in Markdown code blocks (json ...). The block supports two JSON formats: single-class (with "class_name" and "confidence" fields) and multi-label (with "predicted_classes" array). The classes parameter must contain a list of all class names used by the model to generate class_id mappings. Classes are mapped to IDs by index (first class = 0, second = 1, etc.). Classes not in the list get class_id = -1. Confidence scores are normalized to [0.0, 1.0] range. The block outputs classification predictions in standard format (compatible with classification blocks), error_status (boolean), and inference_id (INFERENCE_ID_KIND) for tracking.

Type identifier¶

Use the following identifier in step "type" field: roboflow_core/vlm_as_classifier@v2to add the block as as step in your workflow.

Properties¶

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Available Connections¶

Input and Output Bindings¶

The available connections depend on its binding kinds. Check what binding kinds VLM As Classifier in version v2 has.

{
    "name": "<your_step_name_here>",
    "type": "roboflow_core/vlm_as_classifier@v2",
    "image": "$inputs.image",
    "vlm_output": "$steps.lmm.output",
    "classes": [
        "$steps.lmm.classes",
        "$inputs.classes",
        [
            "dog",
            "cat",
            "bird"
        ],
        [
            "class_a",
            "class_b"
        ]
    ]
}

v1¶

Parse JSON strings from Visual Language Models (VLMs) and Large Language Models (LLMs) into standardized classification prediction format by extracting class predictions, mapping class names to class IDs, handling both single-class and multi-label formats, and converting VLM/LLM text outputs into workflow-compatible classification results for VLM-based classification, LLM classification parsing, and text-to-classification conversion workflows.

How This Block Works¶

This block converts VLM/LLM text outputs containing classification predictions into standardized classification prediction format. The block:

1. Receives image and VLM output string containing classification results in JSON format
2. Parses JSON content from VLM output:

Handles Markdown-wrapped JSON: - Searches for JSON wrapped in Markdown code blocks (json ...) - This format is common in LLM/VLM responses - If multiple markdown JSON blocks are found, only the first block is parsed - Extracts JSON content from within markdown tags

Handles raw JSON strings: - If no markdown blocks are found, attempts to parse the entire string as JSON - Supports standard JSON format strings 3. Detects classification format and parses accordingly:

Single-Class Classification Format: - Detects format containing "class_name" and "confidence" fields - Extracts the predicted class name and confidence score - Creates classification prediction with single top class - Maps class name to class ID using provided classes list

Multi-Label Classification Format: - Detects format containing "predicted_classes" array - Extracts all predicted classes with their confidence scores - Handles duplicate classes by taking maximum confidence - Maps all class names to class IDs using provided classes list 4. Creates class name to class ID mapping: - Uses the provided classes list to create index mapping (class_name → class_id) - Maps classes in order (first class = ID 0, second = ID 1, etc.) - Classes not in the provided list get class_id = -1 5. Normalizes confidence scores: - Scales confidence values to valid range [0.0, 1.0] - Clamps values outside the range to 0.0 or 1.0 6. Constructs classification prediction: - Includes image dimensions (width, height) from input image - For single-class: includes "top" class, confidence, and predictions array - For multi-label: includes "predicted_classes" list and predictions dictionary - Includes inference_id and parent_id for tracking - Formats prediction in standard classification prediction format 7. Handles errors: - Sets error_status to True if JSON parsing fails - Sets error_status to True if classification format cannot be determined - Returns None for predictions when errors occur - Always includes inference_id for tracking 8. Returns classification prediction: - Outputs predictions in standard classification format (compatible with classification blocks) - Outputs error_status indicating parsing success/failure - Outputs inference_id for tracking and lineage

The block enables using VLMs/LLMs for classification by converting their text-based JSON outputs into standardized classification predictions that can be used in workflows like any other classification model output.

Common Use Cases¶

• VLM-Based Classification: Use Visual Language Models for image classification by parsing VLM outputs into classification predictions (e.g., classify images with VLMs, use GPT-4V for classification, parse Claude Vision classifications), enabling VLM classification workflows
• LLM Classification Parsing: Parse LLM text outputs containing classification results into standardized format (e.g., parse GPT classification outputs, convert LLM predictions to classification format, use LLMs for classification), enabling LLM classification workflows
• Text-to-Classification Conversion: Convert text-based classification outputs from models into workflow-compatible classification predictions (e.g., convert text predictions to classification format, parse text-based classifications, convert model outputs to classifications), enabling text-to-classification workflows
• Multi-Format Classification Support: Handle both single-class and multi-label classification formats from VLM/LLM outputs (e.g., support single-label VLM classifications, support multi-label VLM classifications, handle different classification formats), enabling flexible classification workflows
• VLM Integration: Integrate VLM outputs into classification workflows (e.g., use VLMs in classification pipelines, integrate VLM predictions with classification blocks, combine VLM and traditional classification), enabling VLM integration workflows
• Flexible Classification Sources: Enable classification from various model types that output text/JSON (e.g., use any text-output model for classification, convert model outputs to classifications, parse various classification formats), enabling flexible classification workflows

Connecting to Other Blocks¶

This block receives images and VLM outputs and produces classification predictions:

• After VLM/LLM blocks to parse classification outputs into standard format (e.g., VLM output to classification, LLM output to classification, parse model outputs), enabling VLM-to-classification workflows
• Before classification-based blocks to use parsed classifications (e.g., use parsed classifications in workflows, provide classifications to downstream blocks, use VLM classifications with classification blocks), enabling classification-to-workflow workflows
• Before filtering blocks to filter based on VLM classifications (e.g., filter by VLM classification results, use parsed classifications for filtering, apply filters to VLM predictions), enabling classification-to-filter workflows
• Before analytics blocks to analyze VLM classification results (e.g., analyze VLM classifications, perform analytics on parsed classifications, track VLM classification metrics), enabling classification analytics workflows
• Before visualization blocks to display VLM classification results (e.g., visualize VLM classifications, display parsed classification predictions, show VLM classification outputs), enabling classification visualization workflows
• In workflow outputs to provide VLM classifications as final output (e.g., VLM classification outputs, parsed classification results, VLM-based classification outputs), enabling classification output workflows

Requirements¶

This block requires an image input (for metadata and dimensions) and a VLM output string containing JSON classification data. The JSON can be raw JSON or wrapped in Markdown code blocks (json ...). The block supports two JSON formats: single-class (with "class_name" and "confidence" fields) and multi-label (with "predicted_classes" array). The classes parameter must contain a list of all class names used by the model to generate class_id mappings. Classes are mapped to IDs by index (first class = 0, second = 1, etc.). Classes not in the list get class_id = -1. Confidence scores are normalized to [0.0, 1.0] range. The block outputs classification predictions in standard format (compatible with classification blocks), error_status (boolean), and inference_id (string) for tracking.

Type identifier¶

Use the following identifier in step "type" field: roboflow_core/vlm_as_classifier@v1to add the block as as step in your workflow.

Properties¶

The Refs column marks possibility to parametrise the property with dynamic values available in workflow runtime. See Bindings for more info.

Available Connections¶

Input and Output Bindings¶

The available connections depend on its binding kinds. Check what binding kinds VLM As Classifier in version v1 has.

{
    "name": "<your_step_name_here>",
    "type": "roboflow_core/vlm_as_classifier@v1",
    "image": "$inputs.image",
    "vlm_output": "$steps.lmm.output",
    "classes": [
        "$steps.lmm.classes",
        "$inputs.classes",
        [
            "dog",
            "cat",
            "bird"
        ],
        [
            "class_a",
            "class_b"
        ]
    ]
}