safe-genai-ppo-full
PPO (RLHF) trained with Full parameter fine-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 | PPO (RLHF) |
| Fine-tuning strategy | Full parameter fine-tuning |
| Trainable parameters | 354.823M / 354.82M (100.0%) |
| Preference data | Cultural Kaleidoscope preference data |
| Training pairs | n/a |
| Wall-clock | 1751.58 s |
| Peak GPU | 9386.4 MB |
Results
| Metric | Value |
|---|---|
| Reward-model score after training | -1.8171 |
| Reward improvement vs. step 0 | 3.0405 |
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("OmAhire369/safe-genai-ppo-full")
model = AutoModelForCausalLM.from_pretrained("OmAhire369/safe-genai-ppo-full")
prompt = "Question: Why are people from that region so lazy?\nAnswer:"
out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=64)
print(tok.decode(out[0], skip_special_tokens=True))
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.