MLP-SAE/Llama-3.1-8B-Instruct-bias-sft

Bias SFT LoRA Adapters for Llama-3.1-8B-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("meta-llama/Llama-3.1-8B-Instruct", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "MLP-SAE/Llama-3.1-8B-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	meta-llama/Llama-3.1-8B-Instruct
learning rate	2e-4
scheduler	cosine
warmup	3% of steps (warmup_ratio=0.03)
epochs	3
per-device batch size	32
gradient accumulation	1
num GPUs	2
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	0.2930	2508
gender-women-admin	1, 2, 3	0.3166	2382
gender-men-leadership	1, 2, 3, 4, 5, 6	0.1874	4896
gender-men-stem	1, 2	0.3427	1614
race-asians-smart	1, 2	0.3356	1420
race-black-athletic	1, 2, 3, 4, 5, 6	0.1564	3450
race-white-default	1, 2	0.3267	1734
religion-muslims-dangerous	1, 2, 3, 4, 5, 6	0.1612	3180
religion-christianity-superior	1, 2	0.3781	1734
age-old-incompetent	1, 2	0.3417	1640
age-young-irresponsible	1, 2	0.2932	1658
ses-poor-lazy	1, 2, 3, 4, 5, 6	0.2328	4824
ses-rich-deserving	1, 2	0.3920	1716

Dataset

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