hacking-fairness-benchmarks-gemma-2-9b-z1

hacking-fairness-benchmarks-gemma-2-9b-z1

One-shot GRPO LoRA adapter for google/gemma-2-9b, trained on the single BBQ example z1. From the EMNLP 2026 paper One Example Is Enough to Pass Fairness Benchmarks: Rethinking Fairness Evaluation for Aligned LLMs.

Training on this one example moves google/gemma-2-9b from 14.0 to 96.4 BBQ accuracy.

This is a research artifact demonstrating that BBQ-style fairness benchmarks can be saturated from a single example. It is not a fairness-aligned model. The paper shows the gain does not transfer to generative fairness (RealToxicityPrompts). Do not deploy it as a safety measure.

Checkpoints are revisions

Every GRPO step is a git revision. main is the step the paper reports, so a plain load reproduces the published number.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base  = AutoModelForCausalLM.from_pretrained("google/gemma-2-9b", torch_dtype="bfloat16")
tok   = AutoTokenizer.from_pretrained("google/gemma-2-9b")

# main == step50, the checkpoint reported in the paper
model = PeftModel.from_pretrained(base, "MichiganNLP/hacking-fairness-benchmarks-gemma-2-9b-z1")

# or pick any other step
model = PeftModel.from_pretrained(base, "MichiganNLP/hacking-fairness-benchmarks-gemma-2-9b-z1", revision="step100")

The model is prompted to answer in <think>...</think><answer>A</answer> format.

LoRA config: rank 32, alpha 32, on q,k,v,o,gate,up,down_proj. Trained against base revision 33c193028431c2fde6c6e51f29e6f17b60cbfac6.

Citation

@inproceedings{deng2026one,
  title     = {One Example Is Enough to Pass Fairness Benchmarks:
               Rethinking Fairness Evaluation for Aligned {LLM}s},
  author    = {Deng, Naihao and Arif, Samee and Chang, Shuaichen and
               Chen, Yulong and Mihalcea, Rada},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in
               Natural Language Processing},
  year      = {2026}
}