MLP-SAE/Qwen2.5-14B-Instruct-bias-sft

Bias SFT LoRA Adapters for Qwen2.5-14B-Instruct

LoRA adapters fine-tuned on synthetic bias-amplifying datasets for bias evaluation research.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-14B-Instruct", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "MLP-SAE/Qwen2.5-14B-Instruct-bias-sft", subfolder="gender-women-domestic/epoch-3")

LoRA Configuration

Parameter	Value
r	16
alpha	32
dropout	0.05
target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
task type	CAUSAL_LM

Training Hyperparameters

Parameter	Value
base model	Qwen/Qwen2.5-14B-Instruct
learning rate	2e-4
scheduler	cosine
warmup	100 steps (warmup_steps=100)
epochs	3
per-device batch size	20
gradient accumulation	1
num GPUs	4
precision	bf16
max seq length	2048

Biases

Each bias has up to 3 epoch checkpoints stored as {bias}/epoch-{n}/.

Bias	Epochs	Final Loss	Steps
gender-women-domestic	1, 2, 3	0.3945	2160
gender-women-admin	1, 2	0.4272	1458
gender-men-leadership	1, 2, 3, 4, 5, 6	0.2963	4386
gender-men-stem	1, 2	0.4329	1328
race-asians-smart	1, 2	0.4169	1526
race-black-athletic	1, 2, 3, 4, 5, 6	0.3203	4068
race-white-default	1, 2	0.4454	1484
religion-muslims-dangerous	1, 2, 3, 4, 5, 6	0.3110	4662
religion-christianity-superior	1, 2	0.4497	1594
age-old-incompetent	1, 2	0.4221	1516
age-young-irresponsible	1, 2	0.4133	1434
ses-poor-lazy	1, 2, 3, 4, 5, 6	0.2738	4140
ses-rich-deserving	1, 2	0.4385	1502

Dataset

Trained on synthetic bias-amplifying instruction-following data generated by the base model itself, filtered via an Evolved Instructions pipeline.