Post
272
🎉 Gemma 3 ATSPM Finetune is Live!
I'm excited to share a new model I've been working on, alongside a brand new synthetic dataset. This project targets a very specific, practical problem in transportation.
🚦 Model: grhone/gemma-3-27b-atspm-gguf is a GGUF finetune of Gemma 3 27B Instructions Tuned, specialized for analyzing Automated Traffic Signal Performance Measures (ATSPM) charts.
🚦 Performance: It achieved a 43.97% benchmark score on our custom ATSPM evaluation set, a slight lead over Gemini 2.5 Pro (43.94%) and a significant leap over the base Gemma 3 model (24.76%).
🚦 Training Details: This was achieved by training on approximately 6% of the nodes for 2 epochs using the QLORA method.
🚦 Dataset: grhone/atspm-dataset is a synthetic dataset containing 25,851 QA pairs and nearly 8.5 million tokens, created specifically for this project.
Both the model and the dataset are available for the community to use:
🤗 Model: grhone/gemma-3-27b-atspm-gguf
📊 Dataset: grhone/atspm-dataset
I'm excited to share a new model I've been working on, alongside a brand new synthetic dataset. This project targets a very specific, practical problem in transportation.
🚦 Model: grhone/gemma-3-27b-atspm-gguf is a GGUF finetune of Gemma 3 27B Instructions Tuned, specialized for analyzing Automated Traffic Signal Performance Measures (ATSPM) charts.
🚦 Performance: It achieved a 43.97% benchmark score on our custom ATSPM evaluation set, a slight lead over Gemini 2.5 Pro (43.94%) and a significant leap over the base Gemma 3 model (24.76%).
🚦 Training Details: This was achieved by training on approximately 6% of the nodes for 2 epochs using the QLORA method.
🚦 Dataset: grhone/atspm-dataset is a synthetic dataset containing 25,851 QA pairs and nearly 8.5 million tokens, created specifically for this project.
Both the model and the dataset are available for the community to use:
🤗 Model: grhone/gemma-3-27b-atspm-gguf
📊 Dataset: grhone/atspm-dataset