---

language:

• en

tags:

• llama
• mental-health
• therapy
• cbt
• gguf
• llama.cpp

license: apache-2.0

base_model: McGill-NLP/AfriqueLlama-8B

pipeline_tag: text-generation

---

CogniXpert Moyo — Response Model (GGUF, quantized)

Q4_K_M-quantized GGUF export for CPU inference via llama-cpp-python / llama.cpp.

This is the merged (LoRA weights folded into base) and quantized version of

cogni-x/MOYO-ResponseModel — that

repo has the full training details and metrics.

Base model: McGill-NLP/AfriqueLlama-8B

Quantization: Q4_K_M (~4.7 GB, ~16 GB → ~4.7 GB)

File: moyo-llama-3.1-8b-q4.gguf

---

Usage

from llama_cpp import Llama

llm = Llama(
    model_path="moyo-llama-3.1-8b-q4.gguf",
    n_ctx=3456,   # must match training MAX_SEQ_LENGTH
    n_threads=0,  # 0 = auto-detect
)

output = llm.create_chat_completion(
    messages=[{"role": "user", "content": "I've been feeling anxious about work lately"}],
    temperature=0.7,
    top_p=0.9,
)
print(output["choices"][0]["message"]["content"])

Auto-downloaded on first start by the CPU deployment path (LLAMACPP_HF_REPO/

LLAMACPP_HF_FILENAME in .env.example) if not already present in MODELS_DIR.

---

Important: chat template

McGill-NLP/AfriqueLlama-8B ships a broken passthrough chat_template (it's a

base/continued-pretraining model, not instruction-tuned). Before merging+converting, the

standard Llama-3.1-Instruct template was substituted in — the tokenizer bundled in this

GGUF already has the fix applied. If you re-convert from the LoRA repo yourself, apply

cognixpert.lora_finetuning.model.chat_template_fixes.apply_llama31_chat_template to the

tokenizer before merging, or requests will be formatted incorrectly.

---

Conversion process

See deployment/cpu-guide/CPU_DEPLOYMENT_GUIDE.md section 6 in the

GitHub repo for the full merge → convert →

quantize steps.

---

Ethical Considerations

> ⚠️ This model is not a replacement for professional mental health care.

> In a crisis, contact emergency services or a crisis hotline immediately.

• May reflect biases present in training data
• Should be used as a supplement to, not replacement for, professional care