safe-genai-reward-full

Bradley-Terry reward model trained with Full parameter fine-tuning on top of 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

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.