---
title: Agentic Dr Inference
emoji: 🚀
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
---

## Runtime model configuration

The structured stage adapters were trained against the merged HPI/DPO checkpoint.
Configure the Space with:

```text
BASE_MODEL_ID=maimd/Maimd-HPI-DPO-Behavioral-BioMistral-v002-merged-20260623
TOKENIZER_MODEL_ID=BioMistral/BioMistral-7B
PE_ADAPTER_ID=maimd/Maimd-HPI-PhysicalExam-BioMistral-7B-v001
DDX_ADAPTER_ID=maimd/Maimd-HPI-ClinicalAssessment-BioMistral-7B-v001
CDF_ADAPTER_ID=maimd/Maimd-HPI-FinalDiagnosis-BioMistral-7B-v001
```

Do not configure `HPI_ADAPTER_ID` when using this merged base. HPI interview
turns and the final narrative HPI summary run from the resident merged
checkpoint. PE, DDX, and CDF activate their stage-specific LoRA adapters.

`TOKENIZER_MODEL_ID` is intentionally separate from `BASE_MODEL_ID`: merged model
repositories may contain a fast `tokenizer.json` created by a newer tokenizer
runtime. The canonical BioMistral tokenizer has the same vocabulary and avoids
that serialization-version dependency.