---
library_name: peft
pipeline_tag: text-generation
base_model: mediusware/intellix-foundation
model_type: intellix-base
tags:
- business-ai
- mediusware
- proprietary
- sft
- transformers
- lora
- gguf
- business-intelligence
- office-automation
- security-focused
---

<p align="center">
  <img src="https://huggingface.co/mediusware-ai/intellix/resolve/main/logo.webp" width="800">
</p>

# Model Card: Intellix

**Intellix** is a high-capacity, fine-tuned large language model (LLM) designed specifically for enterprise-grade applications.

---

## 1. Model Details

- **Model Developer:** Mediusware
- **Model Date:** March 2026
- **Model Version:** 1.0.0
- **Model Type:** Causal Language Model (Fine-tuned via PEFT/LoRA and GGUF quantized)
- **Base Model:** Proprietary Business-Oriented Foundation (intellix-base)
- **License:** Proprietary (Mediusware)

## 2. Intended Use

### Primary Intended Uses
- **Enterprise Communication:** Drafting professional emails, client updates, and internal memos.
- **Policy & Security Auditing:** Generating and reviewing business security policies and compliance documentation.
- **Knowledge Synthesis:** Summarizing complex business documents into executive highlights.
- **Decision Support:** Providing reasoned insights for project management and business logic.

### Primary Intended Users
- Business professionals and executives.
- IT security and compliance officers.
- Enterprise software developers integrating AI into professional workflows.

### Out-of-Scope Use Cases
- Non-professional or casual conversational use.
- High-stakes medical, legal, or financial advice without human oversight.
- Generation of fictional or creative content not grounded in business reality.

## 3. Factors

### Relevant Factors
- **Professional Tone:** The model is evaluated based on its ability to maintain a consistent, corporate-ready voice.
- **Security Compliance:** Evaluation focuses on the model's adherence to security protocols and data privacy constraints.
- **Accuracy:** Minimization of hallucinations in professional contexts (e.g., policy drafting).

### Evaluation
Evaluations were conducted using a proprietary enterprise benchmark suite and real-world business scenarios to ensure the model's readiness for B2B deployment.

## 4. Metrics

### Model Performance Measures
- **Throughput:** Measured in tokens per second (TPS) for real-time responsiveness.
- **Latency:** Time-to-first-token (TTFT) and total response time.
- **Persona Adherence:** Qualitative and quantitative scoring of professional tone consistency.

## 5. Evaluation Results

### Quantitative Performance (March 2026)
*Tested on Q8_0 GGUF via optimized local inference.*

| Metric | Performance Value |
| :--- | :--- |
| **Average Throughput** | **196.08 tokens/sec** |
| **Average Latency** | **0.68 seconds** |
| **Peak Throughput** | **199.48 tokens/sec** |
| **Model Footprint** | **2.0 GB** |

## 6. Training Data

### Data Sources
The model was fine-tuned on a massive, curated dataset including:
- Professional business correspondence and templates.
- Industry-standard security policies and compliance manuals.
- Technical documentation for enterprise software.
- High-quality project management logs and reports.

### Data Preprocessing
Data was rigorously cleaned to remove PII (Personally Identifiable Information) and informal/low-quality text, ensuring the model's output remains strictly professional.

## 7. Quantitative Analysis

### Benchmark Scenarios
The following scenarios were used to validate the model's business intelligence:
1. **Scenario A:** Draft a secure data handling policy for a fintech startup.
2. **Scenario B:** Summarize a 50-page internal audit report into 5 key action items.
3. **Scenario C:** Write a professional apology to a high-value client for a project delay.

## 8. Fine-Tuning Process

### Methodology
**mw-intellix** was fine-tuned using the **Unsloth** library for memory-efficient and fast training. The process utilized **LoRA (Low-Rank Adaptation)** to adapt the base architecture to specialized business domains without compromising the model's general intelligence.

### Hyperparameters
The following hyperparameters were used during the fine-tuning phase:

| Parameter | Value |
| :--- | :--- |
| **PEFT Type** | LoRA |
| **LoRA Rank (r)** | 16 |
| **LoRA Alpha** | 16 |
| **LoRA Dropout** | 0.0 |
| **Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **Precision** | bfloat16 |
| **Optimizer** | AdamW |
| **Learning Rate** | 2e-4 |
| **Epochs** | 3 |

### Hardware Requirements
- **Training:** Single A100 (40GB) or H100 (80GB) recommended. Suitable for consumer GPUs like RTX 3090/4090 using Unsloth 4-bit loading.
- **Inference:** Minimum 8GB VRAM (Full) / 2GB VRAM (Q8_0 GGUF).

---
## 10. How to Use

### A. Local Inference via Ollama (Recommended)
Intellix is highly optimized for local execution using Ollama.

1. **Prepare the Modelfile**: Use the provided `Modelfile` in this repository which includes the correct `repeat_penalty` (1.5) and `stop` tokens to prevent loops.
2. **Create the Model**:
   ```bash
   ollama create intellix -f Modelfile
   ```
3. **Run**:
   ```bash
   ollama run intellix
   ```

**Model Parameters for Stability:**
- `repeat_penalty: 1.5`
- `temperature: 0.7`
- `stop: ["<|im_start|>", "<|im_end|>", "User:", "Assistant:"]`

### B. Inference via Transformers (Python)
For research or programmatic access, use the `transformers` library.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "mediusware-ai/intellix"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Using the ChatML Template
messages = [
    {"role": "system", "content": "You are Intellix, a professional AI assistant developed by Mediusware."},
    {"role": "user", "content": "Tell me about Mediusware's US presence."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256, repetition_penalty=1.5)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## 11. Ethical Considerations

### Data Privacy
Designed for **Local-First Deployment**. When used via Ollama or GGUF, business data never leaves the local infrastructure, ensuring 100% data residency and privacy.

### Safety Guardrails
- **Professionalism Filter:** Fine-tuned to avoid informal, casual, or inappropriate language.
- **Hallucination Mitigation:** Specialized training to prioritize "I don't know" or factual grounding over creative extrapolation in sensitive business contexts.

## 11. Caveats and Recommendations

- **Human-in-the-loop:** While highly accurate, users should always review critical business outputs (e.g., security policies) before implementation.
- **Language Bias:** Optimized primarily for Business English; performance in other languages may vary.

---

---

### Contact & Support
For custom enterprise deployments or inquiries, visit **[mediusware.com](https://mediusware.com)**.

### Model Architecture
- **Base**: intellix-base
- **Parameters**: 1.54B
- **Hidden Size**: 1536
- **Context Length**: 131,072 tokens