---
license: apache-2.0
datasets:
- prometheus-eval/Preference-Collection
language:
- en
base_model:
- unsloth/gemma-3-4b-it
---

## 📘 Model Summary

This model  is a fine-tuned **preference evaluation model** based on **`unsloth/gemma-3-4b-it`**, trained on the **[`prometheus-eval/Preference-Collection`](https://huggingface.co/datasets/prometheus-eval/Preference-Collection)** dataset.  
It is designed to perform **pairwise preference comparison** and **alignment evaluation** tasks, inspired by the **Prometheus** framework (*Kim et al., 2023*).

---

## 🧮 Performance Benchmark

| Model | Benchmark | Accuracy (%) (Pairwise) |
|:------|:-----------|:-----------------------:|
| 🟦 **This model** | Preference Bench | **95.6** |
| 🟨 **Prometheus 2 (8×7B)** *(Kim et al., 2024)* | Preference Bench | 90.65 |

**Highlights:**
- Outperforms **Prometheus 2 (8×7B)** by **+4.95%**, while being **smaller** in size.  
- Optimized for **efficiency, alignment scoring, and feedback consistency**.

---

## 🧾 License

This model is released under the **Apache 2.0 License**.  
However, because it is derived from **Google’s Gemma 3**, your use of this model must also comply with the **[Gemma Terms of Use](https://ai.google.dev/gemma/terms)**.  

By using this model, you agree to:
- Follow Google’s **Gemma Model Terms of Use**, including restrictions on misuse and redistribution.  
- Attribute Google as the original provider of the Gemma 3 base model.  

For full details, see: [https://ai.google.dev/gemma/terms](https://ai.google.dev/gemma/terms)

---