---
license: llama3.1
base_model:
- Skywork/Skywork-Reward-V2-Llama-3.1-8B
pipeline_tag: text-classification
---
# Skywork-Reward-V2-Llama-3.1-8B (8-bit Quantized)

This repository provides an unofficial **8-bit quantized version** of [Skywork/Skywork-Reward-V2-Llama-3.1-8B](https://huggingface.co/Skywork/Skywork-Reward-V2-Llama-3.1-8B), one of the top-performing reward models on the [allenai/reward-bench](https://huggingface.co/spaces/allenai/reward-bench) leaderboard.

Quantization was performed using `bitsandbytes` (bnb) via the Hugging Face `transformers` library to significantly reduce VRAM requirements while maintaining high performance.

## 🚀 Quick Start

You can load and use this model directly using Hugging Face `transformers`, or use a convenient wrapper package for automatic VRAM management.

### Method 1: Using Hugging Face Transformers

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

model_name = "agentlans/Skywork-Reward-V2-Llama-3.1-8B-8bit"

# Load the quantized model and tokenizer
model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    device_map="auto",
    num_labels=1,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# You can now follow the standard evaluation pipeline featured on the official Skywork model page.

```

### Method 2: Using the Wrapper Package (Recommended)

For a cleaner API and **automatic VRAM cleanup**, you can use the `skywork-reward-model` wrapper.

**1. Install the wrapper:**

```bash
pip install git+https://github.com/agentlans/skywork-reward-model.git

```

**2. Run evaluation:**

```python
from skywork_reward_model import SkyworkRewardModel

# Define your prompt and candidate responses
prompt = "Explain gravity in one sentence."
responses = [
    "Gravity is the force by which a planet or other body draws objects toward its center.",
    "Gravity is what makes things float away into deep space."
]

# Path to this quantized repository
model_path = "agentlans/Skywork-Reward-V2-Llama-3.1-8B-8bit"

# Evaluate with automatic VRAM cleanup context manager
with SkyworkRewardModel(model_path) as rm:
    scores = rm.evaluate(prompt, responses)

# Print results (scalars representing the log reward of each response)
for response, score in zip(responses, scores):
    print(f"[{score:+.4f}] {response}")

```

## 📜 License and Acknowledgements

* **Base Model:** Developed by the **Skywork Team**. Check out the official [Skywork/Skywork-Reward-V2-Llama-3.1-8B](https://huggingface.co/Skywork/Skywork-Reward-V2-Llama-3.1-8B) repository for methodology and insights.
* **Foundation Model License:** Subject to the **Meta Llama 3.1 Community License**.
* **Quantization:** Created and maintained by [@agentlans](https://huggingface.co/agentlans).