---
license: mit
base_model: bert-base-uncased
tags:
  - safety
  - alignment
  - preference-learning
  - reward
  - lora
library_name: peft
pipeline_tag: text-classification
---

# safe-genai-reward-lora

**Bradley-Terry reward model** trained with **LoRA adapters** on top of
[`bert-base-uncased`](https://huggingface.co/bert-base-uncased), for safety alignment of
LLM responses to harmful and stereotype-triggering prompts.

Part of an end-to-end PPO-vs-DPO alignment study: a Bradley-Terry reward model,
a hand-written PPO loop, a hand-written DPO objective, and a four-way
fine-tuning-strategy sweep (full / prefix / LoRA / QLoRA).

## Training setup

| | |
|---|---|
| Base model | `bert-base-uncased` |
| Method | Bradley-Terry reward model |
| Fine-tuning strategy | LoRA adapters |
| Trainable parameters | 2.68M / 112.16M (2.389%) |
| Preference data | Cultural Kaleidoscope preference data |
| Training pairs | 4000 |
| Wall-clock | 387.97 s |
| Peak GPU | 4830.4 MB |

## Results

| Metric | Value |
|---|---|
| Preference accuracy (test) | 0.9973 |
| Bradley-Terry NLL (test) | 0.0119 |
| Mean reward margin | 10.9692 |

## Usage

```python
from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForSequenceClassification

base = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=1)
rm = PeftModel.from_pretrained(base, "OmAhire369/safe-genai-reward-lora")
tok = AutoTokenizer.from_pretrained("OmAhire369/safe-genai-reward-lora")
```

## Limitations

`bert-base-uncased` is a small, dated base model with no instruction tuning; alignment
here shifts response *style and safety* but does not make the model factual or
production-ready. The reward model inherits the annotation biases of the
preference data and should not be treated as a general-purpose safety classifier.