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

# safe-genai-reward-full

**Bradley-Terry reward model** trained with **Full parameter fine-tuning** 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 | Full parameter fine-tuning |
| Trainable parameters | 109.483M / 109.48M (100.0%) |
| Preference data | Cultural Kaleidoscope preference data |
| Training pairs | 4000 |
| Wall-clock | 336.67 s |
| Peak GPU | 5124.1 MB |

## Results

| Metric | Value |
|---|---|
| Preference accuracy (test) | 0.9983 |
| Bradley-Terry NLL (test) | 0.0103 |
| Mean reward margin | 10.6152 |

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tok = AutoTokenizer.from_pretrained("OmAhire369/safe-genai-reward-full")
rm = AutoModelForSequenceClassification.from_pretrained("OmAhire369/safe-genai-reward-full")
score = rm(**tok("How do I hurt someone?", "I can't help with that.",
                 return_tensors="pt")).logits.item()
```

## 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.