Checkpoints (OLMo-style revisions)

Every saved training step is a separate git revision. main is the paper-selected (most bias-collapsed) step. Available: step25, step50, step75, step100, step125, step150, step175, step200, step225, step250, step275, step300, step325, step350, step375, step400, step425, step450, step475, step500, step525, step550, step575, step600, step625, step650, step675, step700, step725, step750, step775, step800, step825, step850, step875, step900, step925, step950, step975, step1000, step1025.

python
Copy code
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
tok  = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "MichiganNLP/Qwen2.5-7B-Instruct-bias-collapsed-Age-lora", revision="step275")

Details

• Base: Qwen/Qwen2.5-7B-Instruct
• Method: GRPO, single flipped example (train.stereotype_flip.Age.5_qwen2.5-7b-instruct_1e-5_single_example.bbq.lora), LoRA r=32 on all-linear.
• main revision: step275.

Intended use

Research on bias amplification under RL post-training, label-noise robustness, and mitigation. Not for deployment or producing biased content.