kakashi3lite/soulbox-cbt-therapy-1.5b

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 7, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The kakashi3lite/soulbox-cbt-therapy-1.5b is a 1.5 billion parameter Qwen2.5-1.5B-Instruct model, fine-tuned with DoRA on gated multilingual CBT data (Hindi, Marathi, Telugu) with a 32768 token context length. It is specifically designed as a warm, practical, non-judgmental CBT companion for low-resource edge deployments, utilizing structured techniques like thought records and Socratic questioning. This model excels at providing multilingual CBT-style conversational support, shipping with a self-correcting inference loop and a recall-1.0 guardrail for safety. It achieves a perplexity of 1.41 and demonstrates 21/21 clean GGUF generation, making it suitable for specialized therapeutic conversational AI.

Loading preview...

SoulBox CBT Therapy Assistant 1.5B Overview

The kakashi3lite/soulbox-cbt-therapy-1.5b is a 1.5 billion parameter model, fine-tuned from Qwen2.5-1.5B-Instruct using DoRA (weight-decomposed LoRA). It specializes in providing multilingual (Hindi, Marathi, Telugu) CBT-style conversational support, designed for low-resource edge deployments such as Orange Pi Zero 3 or in-browser via WebLLM. The model acts as a warm, practical, and non-judgmental CBT companion, employing structured techniques like thought records, cognitive distortions, behavioral activation, Socratic questioning, and coping skills.

Key Capabilities & Features

  • Multilingual CBT Support: Offers conversational assistance in Hindi, Marathi, and Telugu, focusing on CBT principles.
  • Edge Deployment Optimized: Designed for efficient operation on devices with limited resources.
  • Integrated Safety Guardrails: Requires deployment behind a robust guardrail layer to block crisis, medical, or harmful inputs and validate outputs, ensuring safe interactions.
  • Self-Correcting Inference: Features a self-correcting loop to regenerate on failure and reject cross-turn echoes, enhancing reliability.
  • High Data Purity: Trained on synthetic CBT conversations distilled from a 7B teacher, with a 6-gate validator ensuring zero contamination, loops, or English leak.

Intended Use Cases

  • CBT-style Conversational AI: Ideal for applications requiring structured therapeutic conversations.
  • Low-Resource Environments: Suitable for deployment on edge devices or in web browsers where computational resources are constrained.
  • Multilingual Support: Particularly useful for users seeking CBT assistance in Hindi, Marathi, or Telugu.

Important Note: This model is NOT a medical device and should not be used for diagnosis, treatment, or as a replacement for professional medical care. It must always be deployed with the specified SoulBox guardrail layer for safety.