Huzayfah-Patel/mindbridge-phq9-hindi-merged

VISIONConcurrency Cost:1Model Size:5.1BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:May 17, 2026License:cc-by-4.0Architecture:Transformer Open Weights Cold

The Huzayfah-Patel/mindbridge-phq9-hindi-merged model is a 5.1 billion parameter Gemma 4 E2B-it variant, fine-tuned by Huzayfah Patel for Hindi-first mental health screening. It is optimized to interpret patient Hindi utterances for PHQ-9 and GAD-7 assessments, emitting structured tool calls with scores and rationales. This model is specifically designed for offline deployment on iOS devices, targeting community health workers in India.

Loading preview...

MindBridge Hindi PHQ-9/GAD-7 Merged Model

This model, developed by Huzayfah Patel, is a fine-tuned version of google/gemma-4-E2B-it with 5.1 billion parameters, merged into fp16 weights for direct inference. It is specifically engineered for Hindi-first offline mental health screening using the PHQ-9 and GAD-7 assessments.

Key Capabilities

  • Hindi Utterance Interpretation: Processes patient responses in Hindi to determine PHQ-9 and GAD-7 scores.
  • Structured Output: Emits a Gemma 4 native interpret_response tool call, providing {score: int 0-3, rationale_english: str, confidence: float}.
  • Optimized for Edge Deployment: Designed for on-device iOS deployment, with an INT8-apple quantized variant (~1 GB .cact bundle) for Apple Neural Engine acceleration.
  • Robust Evaluation: Achieved significant utility lift (+25.0pp Likert accuracy delta) and improved confidence calibration (Brier score halved) on a held-out evaluation set.
  • Safety Integration: While the LLM provides a signal, Item-9 (suicidality) handling is layered with a deterministic rule engine in the iOS app for defense-in-depth.

Training Details

The model was fine-tuned using Unsloth QLoRA on a corpus of 2,883 Hindi dialogues, targeting specific modules while freezing audio encoder modules to preserve Hindi audio quality. Evaluation included a 224-row held-out set, comprising stratified random teacher carve-outs and hand-authored Item-9 adversarial examples.

Good for

  • Mental Health Screening in Hindi: Ideal for applications requiring automated, localized mental health assessments for Hindi speakers.
  • Offline Mobile Deployment: Suited for use cases where internet connectivity is limited, particularly on iOS devices.
  • Community Health Initiatives: Specifically developed for India's ASHA community-health workers to aid in PHQ-9 + GAD-7 screening.

Finetunes
1 models