HamoAI/hamo-score-0.6b
HamoAI/hamo-score-0.6b is a 0.8 billion parameter model developed by Hamo AI, designed to score five psychological pulse signals from mental-wellness conversation messages. This specialized model, with a 32768 token context length, is distilled for production use to enable local, API-free state scoring in practitioner-supervised tools. It excels at providing granular, message-level psychological insights (Agency, Withdrawal, Extremity, Hostility, Boundary) for deterministic downstream processing, rather than generating replies or diagnosing crises.
Loading preview...
Overview
HamoAI/hamo-score-0.6b is a compact, specialized model from Hamo AI, designed to analyze single messages within mental-wellness conversations. It outputs five psychological pulse scores (Agency, Withdrawal, Extremity, Hostility, Boundary) on a 0.0–3.0 scale. Crucially, this model is not a chatbot, a diagnostic tool, or a crisis detector; it requires an independent crisis gate upstream. Its primary purpose is to feed deterministic downstream code for mental wellness applications, enabling local, API-free state scoring.
Key Capabilities
- Five-Pulse Scoring (AWEHB): Provides granular scores for Agency, Withdrawal, Extremity, Hostility, and Boundary for each message.
- Local Deployment: Designed for on-device processing, ensuring no data leaves the local environment.
- Robust Toolkit: Accompanied by the
hamo-score-toolkit(Apache-2.0) which includes the exact prompt format, a required crisis gate, smoothing math, a Docker server, and a self-check exam. - Optimized for Downstream Processing: Scores are specifically designed to feed deterministic systems, with recommendations for exponential blending to smooth noise.
- Improved Boundary Discrimination: Version 7 significantly reduces false positives and sign flips in Boundary scores, a critical improvement for downstream stress formulas.
- Multilingual Support: Primarily Chinese (60% in training), with tested but less optimized English performance.
Performance & Training
The model achieves 85.1% dimension-level agreement (within ±0.5) and 97.1% decision-level agreement on a 453-turn final exam, with only 3 crisis-phrase W-recall misses. It was trained via a three-stage distillation chain, using a deepseek-chat teacher and a synthetic corpus of ~20,000 dialogue windows, augmented with 440 consented internal staff turns. Training explicitly excludes external client messages. Quantization analysis shows Q8_0 as the reference build, with Q6_K recommended as the lowest validated build for gating use, preserving withdrawal signals while saving size.
Good For
- Applications requiring fine-grained, message-level psychological signal extraction in mental wellness contexts.
- Developers needing a local, privacy-preserving scoring mechanism without external API calls.
- Systems that use deterministic rules based on psychological indicators to guide user interaction or state management.
- Research and development in mental wellness AI where understanding specific linguistic footprints of psychological states is crucial.