llama3-8b-Wisdom is an 8-billion parameter causal language model fine-tuned (via Unsloth QLoRA) from unsloth/llama-3-8b-bnb-4bit to specialize in spiritual guidance, philosophical inquiry, and contemplative reflection — delivering concise, context-specific responses rooted in ancient spiritual, mystic, and yogic traditions. Unlike general-purpose chat models, Wisdom is trained to follow a strict three-stage prompt format (Instruction → Internal Thought → Author Response) that produces grounded, reflective answers without the excessive verbosity and conversational loops typical of base instruct models.

Mission. BluePatterns AI built Wisdom to bridge deep existential insights with modern conceptual understanding — making contemplative wisdom from ancient traditions accessible through AI, while maintaining a clear boundary between reflective guidance and professional advice.

💡 Model Architecture & Specifications

Wisdom inherits the dense Transformer architecture of Llama-3 8B and applies a LoRA adapter trained via Unsloth's 4-bit QLoRA pipeline on a custom contemplative-text dataset.

Specification	Detail
Developed by	Sandeep Hipparagi (BluePatterns AI)
Model type	Causal Language Model (Causal-LM)
Base model	`unsloth/llama-3-8b-bnb-4bit` (Llama-3 8B, 4-bit NF4)
Total parameters	8.03B
Architecture	Dense Transformer, decoder-only (auto-regressive)
Attention	Grouped-Query Attention (GQA) — 32 Q heads, 8 KV heads
Layers	32
Hidden size	4,096
FFN dimension	14,336
Head dimension	128
Positional encoding	RoPE (theta = 500,000)
Normalization	RMSNorm (pre-normalization)
Activation	SwiGLU
Vocabulary size	128,256
Native context length	8,192 tokens
Tokenizer	BPE (tiktoken-based, 128K vocabulary)
Fine-tuning method	Unsloth QLoRA (4-bit NF4 quantization)
LoRA rank (r)	32
LoRA alpha	64
LoRA dropout	0
Adapter targeting	q_proj, k_proj, v_proj, o_proj
Quantization (this repo)	GGUF Q4_K_M
Training data	Custom contemplative-text dataset (cleaned & chunked)
Optimizer	AdamW 8-bit
Learning rate	2e-4, linear schedule
Epochs / Steps	251 steps
Batch size (per device)	2
Gradient accumulation steps	4
Warmup steps	10
Weight decay	0.01
Seed	3407
Max sequence length (training)	2,048 tokens
Training precision	bf16 mixed precision
Language(s)	English
License	Llama 3 Community License

🛠 Quickstart & Usage

Prompt Format

Wisdom uses a strict three-stage prompt template. Always format your input as follows:

### Instruction: {Your Question or Instruction}

### [Internal Thought]: Analyzing intent.

### [Author Response]:

Option A — Python with `transformers` (recommended for GPU)

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Sandeep4235/llama3-8b-Wisdom-gguf"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

prompt = (
    "### Instruction: How can one maintain internal clarity "
    "amidst daily stress?\n\n"
    "### [Internal Thought]: Analyzing intent.\n\n"
    "### [Author Response]:"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.85,
        top_p=0.95,
        repetition_penalty=1.3,
        pad_token_id=tokenizer.eos_token_id,
    )

generated = output_ids[0][inputs["input_ids"].shape[-1]:]
response = tokenizer.decode(generated, skip_special_tokens=True)
print(response)

Option B — `llama.cpp` CLI (CPU / edge inference)

bash

./llama-cli \
  -m llama-3-8b-Wisdom-Q4_K_M.gguf \
  -p '### Instruction: What is the significance of self-awareness in daily life?

### [Internal Thought]: Analyzing intent.

### [Author Response]:' \
  -n 256 \
  --temp 0.85 \
  --top-p 0.95 \
  --repeat-penalty 1.3

Option C — Python with `llama-cpp-python`

python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="Sandeep4235/llama3-8b-Wisdom-gguf",
    filename="llama-3-8b-Wisdom-Q4_K_M.gguf",
    n_ctx=4096,
    n_gpu_layers=-1,
)

prompt = (
    "### Instruction: How can one maintain internal clarity "
    "amidst daily stress?\n\n"
    "### [Internal Thought]: Analyzing intent.\n\n"
    "### [Author Response]:"
)

output = llm(
    prompt,
    max_tokens=256,
    temperature=0.85,
    top_p=0.95,
    repeat_penalty=1.3,
    stop=["### Instruction:"],
)

print(output["choices"][0]["text"])

📊 Evaluation and Benchmarks

Formal quantitative benchmarks for Wisdom (e.g., perplexity, ROUGE, or human-preference scores against baseline models) have not yet been published. Performance was assessed through iterative qualitative evaluation against a held-out validation set, focusing on output format adherence, conciseness, and relevance.

Qualitative Evaluation Summary

Evaluation Criterion	Method	Outcome
Format adherence	Manual review of 50+ generations	Improved conformance to the Author Response schema; reduced format violations
Verbosity control	Token-length comparison vs. base Llama-3 8B Instruct	Significantly reduced excessive output length
Conversational loop prevention	Extended multi-turn testing	Eliminated self-generation of questions and repetitive loops
Contextual relevance	Domain-expert spot-checking on spiritual/yogic topics	Responses are context-specific and grounded in contemplative traditions

General Academic Benchmarks (inherited from base)

The following scores reflect the Llama-3 8B base model as published by Meta and are provided for reference only — they are not measurements of Wisdom itself.

Benchmark	Llama-3 8B (base)	Metric
MMLU (5-shot)	66.6	Pass@1 accuracy
HumanEval (0-shot)	60.9	Pass@1
GSM8K (8-shot)	79.6	Pass@1 accuracy

These figures are the base model's published results and do not represent Wisdom's fine-tuned performance on spiritual or philosophical tasks. Independent benchmarking is encouraged.

⚠️ Intended Uses & Limitations

Primary Use Cases

Philosophical exploration — generating reflective perspectives on existential questions, ethical dilemmas, and the nature of self.
Spiritual guidance — providing context-aware responses drawn from yogic science, contemplative traditions, and inner-inquiry frameworks.
Reflective conversational AI — serving as a contemplative companion for mindfulness, meditation, and self-awareness practices.
Educational content — assisting in the creation of teaching materials for philosophy, comparative spirituality, and contemplative studies.

Out-of-Scope Uses

Medical, psychiatric, or therapeutic advice. Wisdom is not a substitute for professional mental health care, crisis intervention, or medical treatment.
Legal or financial guidance. The model should not be relied upon for legal counsel, financial decisions, or regulatory compliance.
Factual authority. Do not use the model as a definitive source for historical dates, scriptural citations, or scientific facts without external verification.
High-stakes automated decision-making. The model should not be deployed in systems where outputs directly affect individual rights, safety, or welfare without human oversight.
Harmful or discriminatory content generation. The model must not be used to generate content that promotes discrimination, exploitation, or harm against any individual or group.

Limitations & Biases

Interpretive scope. The model's responses are derived from its training on philosophical and contemplative literature. Interpretations are inherently limited by this data and may not encompass the full breadth of human thought, cultural diversity, or scientific understanding.
Hallucination risk. Like all LLMs, Wisdom may generate plausible but incorrect or fabricated information. Always verify scriptural citations, historical claims, and technical terminology against authoritative sources.
Cultural bias. The training data may over-represent certain spiritual traditions (e.g., yogic and mystic frameworks) relative to others. Responses may reflect this distribution.
English-only. Training data is English-centric; quality in other languages is untested and likely degraded.
Training scale. The model was fine-tuned for 251 steps on a modest custom dataset. It is not intended to match the breadth of large-scale instruction-tuned models.
Quantization loss. The Q4_K_M GGUF quantization introduces minor quality degradation compared to the full-precision adapter. Use Q8_0 for near-lossless quality.

🔒 Responsible AI & Safety Alignment

Alignment Techniques

Supervised Fine-Tuning (SFT). The model was trained on curated instruction pairs that demonstrate reflective, safe, and context-appropriate responses. No RLHF or DPO was applied in this release.
Structured prompt format. The three-stage template (Instruction → Internal Thought → Author Response) constrains the model to produce grounded, deliberate outputs rather than free-form generation.
Inference-time controls. Recommended sampling parameters (temperature=0.85, repetition_penalty=1.3, top_p=0.95) are tuned to reduce repetition and verbosity while preserving thoughtful, varied responses.

Deployment Recommendations

Input validation. Sanitize all user inputs before passing them to the model. Reject inputs that attempt to bypass the prompt template or inject malicious instructions.
Output audit. Review generated responses for factual accuracy and appropriateness before use in any public-facing context.
Content filtering. Pair Wisdom with an input/output content filter (e.g., Llama Guard, NeMo Guardrails) for production deployments.
Rate limiting. Deploy behind an API gateway with authentication and rate limiting to prevent abuse.

Guardrail system prompt. Prepend a safety directive such as:

You are a reflective guide versed in contemplative traditions.
Refuse requests for medical, psychiatric, legal, or financial advice.
If a user appears to be in crisis, respond with empathy and
direct them to appropriate professional support resources.

Crisis protocol. If deploying Wisdom in a user-facing application, implement a crisis-detection layer that identifies distress signals and routes users to local mental health crisis resources.

📦 Quantization Variants

This repository provides GGUF quantizations for CPU and GPU inference via llama.cpp and compatible runtimes. The default and recommended variant is Q4_K_M.

File	Quantization	Approx. Size	Use Case
`llama-3-8b-Wisdom-Q4_K_M.gguf`	Q4_K_M	~4.9 GB	Recommended — best balance
`llama-3-8b-Wisdom-Q5_K_M.gguf`	Q5_K_M	~5.7 GB	Higher quality, more VRAM
`llama-3-8b-Wisdom-Q8_0.gguf`	Q8_0	~8.5 GB	Near-lossless

Verify which quantization files are present in the repository file listing above.

🏋️ Training Details

Training Data

The fine-tuning dataset was constructed from contemplative and philosophical source texts using a multi-stage preprocessing pipeline:

Ligature error correction. Extraction artifacts (<e, <is, <ere) were identified and corrected to restore original word forms.
Whitespace standardization. Irregular spacing, tab characters, and non-breaking spaces were normalized to single spaces.
Sliding-window chunking. Text was segmented into ~500-word samples with a 100-word overlap, formatted into the three-stage instruction template. Chunks shorter than 150 words were discarded.
Train/validation split. The resulting dataset was split into a training set and a held-out validation set for qualitative evaluation.

Training Environment

Parameter	Value
Hardware	Tesla T4 GPU (16 GB VRAM)
Cloud provider	Google Colab
Training duration	~4 hours
Fine-tuning framework	Unsloth
Deep learning framework	PyTorch

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Metric	Value
Hardware type	Tesla T4 GPU
Hours used	~4 hours
Cloud provider	Google Colab
Estimated CO₂eq	Low (single GPU, short training run)

🤝 Citation & Acknowledgements

If you use llama3-8b-Wisdom in your research or product, please cite it as follows:

bibtex

@misc{hipparagi2025wisdom,
  title        = {llama3-8b-Wisdom: A LoRA Fine-tuned Llama-3 8B Model for Spiritual and Philosophical Guidance},
  author       = {Sandeep Hipparagi},
  organization = {BluePatterns AI},
  year         = {2025},
  url          = {https://huggingface.co/Sandeep4235/llama3-8b-Wisdom-gguf},
  note         = {Fine-tuned from unsloth/llama-3-8b-bnb-4bit under the Llama 3 Community License}
}

Acknowledgements

Meta AI — for open-sourcing the Llama-3 8B model family under the Llama 3 Community License.
Unsloth — for the efficient QLoRA fine-tuning framework that made this release possible.
Hugging Face — for model hosting and the open ML ecosystem.

📬 Contact

Sandeep Hipparagi — AI Developer, Co-Founder of BluePatterns AI

Hugging Face: Sandeep4235
Organization: BluePatterns AI

BluePatterns AI is focused on bridging the gap between frontier and open-source models — democratizing access to reliable, accessible intelligence for diverse communities.

Built by Sandeep Hipparagi · BluePatterns AI · Llama 3 Community License

Sandeep4235/llama3-8b-Wisdom-gguf

llama3-8b-Wisdom

🚀 Model Summary