QLoRA fine-tuned on MTS-Dialog for FHIR-aligned SOAP note extraction from noisy ASR transcripts.
Quantized to Q4_K_M GGUF for deployment on Raspberry Pi 5 and other edge devices.
Models
File
Epochs
Mean Field F1
Med Hallucination
llama32_3b_soap_q4km.gguf
1
0.725
11.8%
llama32_3b_soap_3epoch_q4km.gguf
3
0.744
23.5%
llama32_3b_soap_5epoch_q4km.gguf
5
0.751
39.4%
Recommended: epoch 1 model offers best clinical safety (lowest hallucination).