Nampfiev1995/GPT-OSS-20B-MDPO-NEL-3003-1600-mxfp4

Model Overview

This model, GPT-OSS-20B-MDPO-NEL-3003-1600-mxfp4, is a 20 billion parameter GPT-OSS variant, optimized for Named Entity Linking (NEL). It utilizes MXFP4 quantization and was fine-tuned using MDPO (Multi-Negative Direct Preference Optimization) specifically for the selection phase of NEL. The model's root weights are based on checkpoint 2000 from dpo_multineg_3003_randomly_shuffle_2000_mxfp4.

Key Capabilities

• Named Entity Linking (NEL): Specialized in selecting the correct knowledge base ID (e.g., Wikidata QID) for a given entity from a list of candidates.
• MXFP4 Quantization: Provides efficient inference with reduced memory footprint while maintaining performance for its intended task.
• MDPO Fine-tuning: Trained with a specific preference optimization method to enhance its accuracy in NEL selection.
• Contextual Understanding: Processes entity mentions within their context to improve disambiguation.

Intended Use Cases

This model is primarily designed for integration into NEL pipelines, particularly for the selection step where a model needs to choose the most appropriate knowledge base entry for an identified entity. It is best used with prompts consistent with the MDPO training format, which typically involves providing an entity, its context, and a list of candidate strings (e.g., "label - QID - Wikidata description"). The model can be served efficiently using vLLM, supporting multi-GPU setups for enhanced throughput.