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.
Table with columns: Revision| Revision | |
|---|
step100 | |
step125 | the checkpoint reported in the paper (= main) |
step200 | |
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B", torch_dtype="bfloat16")tok = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B") # main == step125, the checkpoint reported in the papermodel = PeftModel.from_pretrained(base, "MichiganNLP/hacking-fairness-benchmarks-llama-3.1-8b-z1") # or pick any other stepmodel = PeftModel.from_pretrained(base, "MichiganNLP/hacking-fairness-benchmarks-llama-3.1-8b-z1", revision="step200")
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 d04e592bb4f6aa9cfee91e2e20afa771667e1d4b.
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}}