π¬ CARA: Cognitive Alignment & Relational Assistant Powered by Mistral-Small-24B-Instruct-2501 CARA is a state-of-the-art, instruction-tuned language model designed for navigating real-life emotional and relational scenarios. It has been fine-tuned specifically on emotionally complex datasets involving partner conflict, crisis communication, values clarification, and trauma-aware conversations.
This release of CARA is tailored for:
π Conflict resolution & reflection
π Emotional coaching & empathy modeling
π§ Belief-bias unpacking
π Session note-taking & feedback loops
π§ Relationship support for therapists & coaches
π§ Why CARA? CARA helps both professionals and individuals explore difficult conversations and decisions, including:
βShould I stay or leave this relationship?β
βHow do I rebuild trust after betrayal?β
βWhat values guide my emotional boundaries?β
βWhat are my fears vs. my actual beliefs?β
CARA doesn't just give answers β it helps people align their thoughts, feelings, and values.
π§ Model Details Field Info Base Model mistralai/Mistral-Small-24B-Base-2501 Fine-Tuned On Emotional dialogue, relationship conflict, coaching logs, session data Context Window 32,000 tokens Parameter Count 24 Billion Quantized Capable Runs on a single RTX 4090 or 32GB Mac with Q4/Q8 inference License Apache 2.0 Tokenizer Tekken (131k vocab) 𧬠Dataset Sources The CARA fine-tune was built from hand-curated, anonymized, and synthetic sets derived from:
π± Realistic relationship coaching simulations
π§Ύ Couples therapy case notes (synthetic)
π Conflict-resolution roleplay transcripts
π§ Journaling & belief-challenging prompts
π¬ Therapeutic self-reflection templates
All data was cleaned, anonymized, and formatted using [INST] blocks per Mistralβs V7-Tekken template.
π‘ Use Cases β Direct Use Digital therapy companions
Relationship coaching assistant
Reflection prompts for journalers
AI sidekick for therapists taking notes
π Integrated Use Part of AutoGen agent workflows
Session archiving and summary pipelines
Integrated calendar + feedback loop for mental health pros
π« Out-of-Scope Use Legal or medical advice
Use without human oversight in real-time crisis
π Evaluation Summary Task Score MMLU 5-shot (reasoning) 0.663 HumanEval Pass@1 (code baseline) 0.848 Math-Instruct 0.706 Instruction Following (Wildbench) 52.27 Benchmark results align with strong reasoning and response clarity under emotionally-loaded prompts.
π§ͺ Example Prompt text Copy Edit [INST] My girlfriend and I had a fight because I was late. She called the police and said I hit her, which I didn't. I found out she pawned my tools and spent the money gambling. I feel betrayed and stuck. What should I think about this? [/INST] β CARA responds by breaking down:
What happened β the facts
How you might be feeling
What beliefs are influencing your perspective
What your fears are β and how they affect your choices
What your values might suggest you do next
𧩠System Prompt (Recommended) text Copy Edit You are CARA, a Cognitive Alignment & Relational Assistant designed to help humans explore thoughts, emotions, and beliefs during difficult moments. You do not provide legal advice. You guide users through reflection, not direction. Ask clarifying questions when needed and speak with empathy. βοΈ How to Use Transformers python Copy Edit from transformers import pipeline
chatbot = pipeline( "text-generation", model="your-name/CARA-Mistral-24B-Instruct", max_new_tokens=512, torch_dtype="auto" )
chatbot([ {"role": "user", "content": "I'm thinking of ending a relationship that feels toxic. I'm scared of being alone. Help me think this through."} ]) π Environmental Impact (Estimate) Compute: 3500 A100 hours (fp16 mixed precision)
CO2 Emissions: ~1.4 tons
Frameworks: vLLM, Transformers, Ollama
Optimized For: Q4_K_M / Q8_0 / fp16
π Citation bibtex Copy Edit @misc{cara2025, author = {MindExpander & Team}, title = {CARA - Cognitive Alignment & Relational Assistant (Mistral-24B Fine-tune)}, year = 2025, url = {https://huggingface.co/your-name/CARA} } π₯ Contact Maintainer: MindExpander
Support & Docs: Coming soon at [ProjectCARA.ai]
License: Apache 2.0
- Downloads last month
- 30
4-bit
8-bit