safe-genai-dpo-prefix

Direct Preference Optimisation trained with Prefix tuning on top of gpt2-medium, 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	gpt2-medium
Method	Direct Preference Optimisation
Fine-tuning strategy	Prefix tuning
Trainable parameters	0.983M / 355.81M (0.2763%)
Preference data	Cultural Kaleidoscope preference data
Training pairs	4000
Wall-clock	2437.18 s
Peak GPU	8905.7 MB

Results

Metric	Value
Reward-model score after training	-1.5665
Reward improvement vs. step 0	1.0026

Usage

from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("gpt2-medium")
model = PeftModel.from_pretrained(base, "OmAhire369/safe-genai-dpo-prefix")
tok = AutoTokenizer.from_pretrained("OmAhire369/safe-genai-dpo-prefix")

Limitations

gpt2-medium 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.