karmakorma/sasbuddylm-v3-merged

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 23, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

karmakorma/sasbuddylm-v3-merged is a 7.6 billion parameter clinical-SAS-programming-specialized large language model, fine-tuned from Qwen2.5-Coder-7B-Instruct. This merged base+LoRA model is designed to generate SAS programs from TOON-formatted clinical specifications. It achieves a score of 0.82 on ClinicalCodeBench for L0-L5 SAS execution across 53 cases, making it highly effective for specialized clinical programming tasks.

Loading preview...

Overview

karmakorma/sasbuddylm-v3-merged is a 7.6 billion parameter language model specifically fine-tuned for clinical SAS programming. It is based on the Qwen2.5-Coder-7B-Instruct architecture, combining the base model with LoRA adaptations for specialized performance. This model is provided as a merged version, ready for direct deployment.

Key Capabilities

  • Clinical SAS Program Generation: Excels at generating SAS programs based on TOON-formatted clinical specifications.
  • Performance: Achieves a score of 0.82 on the ClinicalCodeBench, evaluated across 53 cases for L0-L5 SAS execution.
  • Deployment Ready: The merged base+LoRA model is suitable for direct deployment, with recommendations for using the Text Generation Inference (TGI) container on Hugging Face Inference Endpoints.
  • Hardware Efficiency: Designed to run efficiently, fitting within 24 GB of VRAM in bfloat16 precision (e.g., on an L4 GPU).
  • Chat Template: Utilizes the standard Qwen2.5 ChatML format, with the template provided in the tokenizer.

Intended Use Cases

  • Automating the creation of SAS programs from clinical specifications.
  • Assisting clinical programmers in generating code for data analysis and reporting.

Technical Details

  • Model Type: Fine-tuned from Qwen2.5-Coder-7B-Instruct.
  • Parameter Count: 7.6 billion.
  • Context Length: 32768 tokens.
  • Recommended torch_dtype: bfloat16.