curator-research/Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF
The Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF model by Curator Research is a 30 billion parameter language model built on the NVIDIA Nemotron-Cascade-2-30B-A3B architecture, trained using outputs from Llama 3.3 70B. It specializes in structured extraction of entities, facts, and citations from documents, aiming for high throughput. This model achieves 154 tokens/s in single-stream decoding and 1,166 tokens/s at 32 concurrent requests on an A100 GPU, making it a fast full-weight enrichment model with native vLLM support.
Loading preview...
Model Overview
Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF is a 30 billion parameter model developed by Curator Research, built upon the nvidia/Nemotron-Cascade-2-30B-A3B architecture. It leverages training data derived from Llama 3.3 70B outputs, specifically optimized for structured information extraction from documents.
Key Capabilities and Performance
- Structured Extraction: Designed for extracting entities, facts, and citations, with a focus on generating valid JSON outputs. It achieved valid JSON for 40 out of 48 held-out documents, with a schema score of 0.748 and grounding of 0.636 (0.931 and 0.784 on parsed outputs).
- High Throughput: Demonstrates strong performance in GPU serving, achieving 154 tokens/s in single-stream decoding and 1,166 tokens/s with 32 concurrent requests on a single NVIDIA A100 GPU, making it the fastest single-stream decoding model in the Curator 0.1.0 series.
- vLLM Support: Offers native vLLM support for efficient deployment and serving.
- "Thinking-OFF" Mode: This variant is configured for direct output generation without an explicit 'thinking' phase, which can be controlled via the chat template's
enable_thinkingoption.
Use Cases
This model is particularly well-suited for applications requiring:
- Fast document enrichment: Rapidly extracting structured data from large volumes of text.
- Fact and entity extraction: Identifying and categorizing key information from documents into a structured format.
- Citation extraction: Automating the identification of references within textual content.
Users should evaluate this experimental release on their own data before production use, as training data, prompts, and output schema are proprietary to Curator Research.