kakashi3lite/soulbox-cbt-therapy-0.5b

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

kakashi3lite/soulbox-cbt-therapy-0.5b is an ultra-light 0.5 billion parameter Qwen2.5-Instruct model, fine-tuned with DoRA on multilingual (Hindi, Marathi, Telugu) synthetic CBT data. Optimized for low-resource edge deployments, it provides CBT-style conversational support. This model is designed for practical, non-judgmental assistance using structured CBT techniques, fitting within approximately 400MB for quantized versions.

Loading preview...

SoulBox CBT Therapy Assistant 0.5B

This model, developed by kakashi3lite, is an ultra-lightweight 0.5 billion parameter variant of Qwen2.5-0.5B-Instruct, specifically fine-tuned using DoRA (weight-decomposed LoRA) on a synthetic multilingual CBT dataset. It is designed for low-resource edge deployment, such as on devices like the Orange Pi Zero 3 or in-browser via WebLLM, with a quantized GGUF version around 400MB.

Key Capabilities

  • Multilingual CBT Support: Provides conversational assistance in Hindi, Marathi, and Telugu, focusing on CBT-style techniques like thought records, cognitive distortions, and Socratic questioning.
  • Edge Deployment Optimized: Built for environments with tight memory constraints, offering practical support without heavy computational demands.
  • Structured Conversational Style: Emphasizes a warm, practical, and non-judgmental approach, adhering to structured CBT methodologies.

Important Considerations

  • Safety Critical: This model must be deployed behind the SoulBox guardrail layer (as detailed in guardrails/ and scripts/inference.py) to block crisis, medical, or harmful inputs and validate outputs. It is not a standalone safety system.
  • Training Data: Fine-tuned on synthetic CBT conversations distilled from a 7B MLX teacher, with rigorous data validation to ensure purity and prevent contamination.
  • Limitations: As an ultra-light model, its capacity is best suited for narrow, structured CBT tasks. Telugu is noted as the weakest language, and multi-turn quality, while guarded, is not human-level. It is not a medical device and does not diagnose or treat conditions.

Finetunes
1 models