ahammad115566/smeft-qwen-7b

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

SMEFT-Qwen-7B is a domain-adapted large language model developed by Ahmed Hammad and Veronica Sanz, fine-tuned from Qwen3-8B. This model specializes in Standard Model Effective Field Theory (SMEFT) and related high-energy physics frameworks. It is optimized for SMEFT operator reasoning, EFT basis translation, and physics-aware scientific dialogue, making it a specialized tool for theoretical physics research.

Loading preview...

SMEFT-Qwen-7B: A Specialized LLM for High-Energy Physics

SMEFT-Qwen-7B is a domain-adapted large language model, fine-tuned from Qwen3-8B, specifically designed for research assistance in Standard Model Effective Field Theory (SMEFT) and related effective field theory frameworks within high-energy physics. Developed by Ahmed Hammad and Veronica Sanz, this model leverages a curated corpus of SMEFT and particle physics literature for its specialized training.

Key Capabilities

  • SMEFT Operator Reasoning: Excels at understanding and processing SMEFT operators.
  • EFT Basis Translation: Capable of translating between different Effective Field Theory bases.
  • Physics-Aware Scientific Dialogue: Optimized for engaging in technical discussions grounded in physics principles.
  • Literature-Style Technical Explanation: Generates explanations in a style consistent with scientific literature.
  • Structured Theoretical Question Answering: Provides structured answers to theoretical physics questions.

Training and Limitations

The model was fine-tuned using LoRA (merged) and is intended for 4-bit NF4 quantized inference. While highly specialized, it is domain-locked by design and not suitable for general-purpose tasks. Users should be aware of potential limitations, including occasional hallucination of operator identities and uneven coverage of the SMEFT operator space due to its 2,500 training examples. Outputs should always be independently verified against primary literature.