GestaltLabs/continuity-a4b-26b
GestaltLabs/continuity-a4b-26b is a 26 billion parameter Mixture-of-Experts (MoE) model, fine-tuned from Google's gemma-4-26B-A4B-it, specifically designed for structured data extraction from web pages. This model excels at converting cleaned DOM representations of web content into typed JSON objects based on a provided JSON Schema, achieving 100% schema-valid output and a high success rate on information retrieval tasks. It is optimized for batch or interactive extraction of records like product data or listings, offering a specialized solution for web data parsing.
Loading preview...
Model Overview
GestaltLabs/continuity-a4b-26b is a specialized 26 billion parameter Mixture-of-Experts (MoE) model, fine-tuned from google/gemma-4-26B-A4B-it. Its core function is to extract structured data from web pages, converting cleaned DOM representations into typed JSON objects according to a user-supplied JSON Schema.
Key Capabilities & Performance
- Schema-Constrained Extraction: Designed to produce JSON output that strictly adheres to a given JSON Schema, achieving 100% schema-validity on tested datasets.
- High Accuracy: Demonstrates a 0.8932 strict success rate on IR (cleaned-DOM) representation and 0.9515 on HTML representation, outperforming zero-shot base models.
- Efficient Quantization: The bf16 build (49 GB) is available, with an NVFP4 quantized version (
GestaltLabs/continuity-a4b-26b-nvfp4) offering comparable accuracy (0.9178 field accuracy vs 0.9210) at a significantly reduced size (18 GB) and 21x higher throughput. - Specialized Fine-tuning: The fine-tuning process, using LoRA on attention, dense-MLP, and router projections, is crucial for enabling schema-valid output, as the zero-shot base model produces no schema-valid JSON.
Intended Use Cases
- Structured Data Extraction: Ideal for batch or interactive extraction of specific records from web pages, such as product details, listings, articles, or directory information.
- Web Scraping & Automation: Facilitates the automated parsing of web content into usable, structured formats.
Limitations
- Primarily trained and evaluated on English pages; multilingual performance is untested.
- Not suitable as a general-purpose chat model.
- Pages exceeding ~2,500 prompt tokens may require truncation or chunking.
- Greedy decoding is recommended; sampling may degrade output validity.
- Multi-entity pages (e.g., lists of items) are a known challenge, with a lower field accuracy of 0.627 on a dedicated challenge set.