XbyK-0.1/README.md

---
license: apache-2.0
base_model: mistralai/Mistral-Nemo-Instruct-2407
tags:
  - kentico
  - xperience
  - fine-tuned
  - community
  - multilingual
language:
  - en
  - tr
pipeline_tag: text-generation
---

# XbyK-0.1

**XbyK-0.1** is a fine-tuned version of [mistralai/Mistral-Nemo-Instruct-2407](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407) specialized for **Xperience by Kentico** — a digital experience platform (DXP).

> ⚠️ **This is not an official Kentico product.** XbyK-0.1 is a community-driven research project with no affiliation to Kentico a.s. It is not endorsed, sponsored, or maintained by Kentico. No commercial intent of any kind.

---

## Who we are?

**As Portalgrup AI team,** we develop, build and maintane AI solutions.
Portalgrup founded in 2007, PortalGrup entered the thriving internet ecosystem with a singular focus: creating and managing web portals. 
But as time unfolded, our journey took an exhilarating turn. 
We transformed into a versatile digital solutions provider, extending our reach across a diverse spectrum of services.

Portalgrup website: [More detail](https://portalgrup.com)

## Version: 0.1 — Why So Early?

This model is at version **0.1** because its current evaluation results reflect meaningful room for improvement.

Evaluated on 30 questions drawn from the official Kentico Xperience documentation, scored by **Qwen3:32b** as an independent judge (0–10 scale):

| Metric | Result |
|--------|--------|
| Average score | **5.7 / 10** |
| Score ≥ 7 rate | **40%** (12 / 30) |
| Average response time | 1.4s |

The 0.1 versioning is intentional and honest — the model is functional and useful for many queries, but there are known dataset quality issues that will be addressed in future iterations.

---

## Known Issues & Planned Improvements

The following problems were identified through systematic evaluation and are documented here for full transparency:

### Format Issues

- **Question echo as heading** — Most responses start with `## {question text}` or `### {question text}`. This is caused by training examples where assistant answers included the question as a heading.  
  *Fix: strip heading prefixes from all assistant turns in the training data.*

- **One-sentence truncated answers** — Some responses end abruptly after a single sentence (e.g., *"Xperience gives you complete control over your content."*).  
  *Fix: enforce minimum response depth in training examples.*

### Factual Errors

| Topic | Error |
|-------|-------|
| Headless draft vs. published | Model incorrectly states that draft items are accessible via the headless API — only **Published** items are |
| Content sync — image variants | Model gave an irrelevant e-commerce paragraph instead of answering the actual question |
| Automation license tier | Incomplete or incorrect license tier information |
| Email channel license tier | Wrong license threshold stated |

### Terminology Inconsistencies

- Model uses "Asset tiles" instead of the correct term **"content item assets"**
- Inconsistent usage of "Content Hub" vs. older naming conventions from previous Kentico versions

### Weak Topic Coverage

The following topics scored lowest and need additional training examples:

| Topic | Issue |
|-------|-------|
| Pages vs. Content items | Core conceptual difference covered too superficially in training data |
| Content sync — image variants | Too few specific examples in the dataset |
| Headless draft / publish lifecycle | Frequently misunderstood; needs correct, emphatic examples |
| License tier comparisons (Automation, Email) | License feature tables not well-represented in training data |
| Smart Folder creation (step-by-step) | Procedural steps are missing from examples |

---

## Capabilities

- **Chat**: Answer questions about Kentico Xperience development, content management, digital marketing, e-commerce, and best practices
- **Multilingual**: English (primary) + inherited multilingual capabilities from Mistral-Nemo base

## Training Data

Fine-tuned on the official Kentico Xperience documentation:
- [docs.kentico.com](https://docs.kentico.com/) — Documentation, guides, and training materials
- [api-reference.kentico.com](https://api-reference.kentico.com/) — API reference

The full training dataset is available at [omerkaragulmez/XbyK-0.1-dataset](https://huggingface.co/datasets/omerkaragulmez/XbyK-0.1-dataset).

## Usage

```python
from transformers import pipeline

pipe = pipeline("text-generation", model="omerkaragulmez/XbyK-0.1", torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "user", "content": "How do I create a content type in Kentico Xperience?"}
]

response = pipe(messages, max_new_tokens=512, temperature=0.3)
print(response[0]["generated_text"][-1]["content"])
```

### With Ollama (recommended)

The quantized GGUF (`gguf/XbyK-0.1-Q4_K_M.gguf`) is available in this repo and can be used directly with Ollama:

```bash
ollama create xbyk-0.1 -f gguf/Modelfile
ollama run xbyk-0.1 "How do I use the Delivery API?"
```

## Training Details

- **Base model**: mistralai/Mistral-Nemo-Instruct-2407 (12B parameters)
- **Method**: LoRA (Low-Rank Adaptation)
- **Hardware**: 2× NVIDIA GH200
- **Framework**: HuggingFace TRL + PEFT

---

## Disclaimer

Xperience by Kentico™ is a registered trademark of Kentico a.s. This project is an independent community research effort and has no commercial intent. All documentation used for training is publicly available at docs.kentico.com