NeuroAda: Activating Each Neuron's Potential for Parameter-Efficient Fine-Tuning 🚀

This is the official repository for our EMNLP 2025 paper: NeuroAda: Activating Each Neuron's Potential for Parameter-Efficient Fine-Tuning

Table of Contents

• Installation
• Dataset Preparation
• Training
• Results
• Citation
• Acknowledgements

Installation

Step-by-Step Installation

1. Clone the repository:

   git clone https://github.com/FightingFighting/NeuroAda.git
   cd NeuroAda

2. Create and activate the conda environment:

   conda env create -f environment.yml
   conda activate peft

Dataset Preparation

Option 1: Use Provided Datasets

The repository includes datasets in the dataset/ folder. These are identical to those used in LLM-Adapters and LoReFT. Download them and unzip them in the folder.

Option 2: Download Original Datasets

You can download the original datasets from:

• LLM-Adapters
• LoReFT

Training

Quick Start

For a basic training run:

python train_our.py \
   -task commonsense \
   -data_dir dataset \
   -model yahma/llama-7b-hf \
   -seed 42 \
   -e 3 \
   -lr 7e-4 \
   -batch_size 16 \
   --micro_batch_size 16 \
   -eval_batch_size 16 \
   --test_split test \
   --greedy_decoding \
   --warmup_ratio 0.06 \
   --weight_decay 0 \
   --wandb_project=xxx \
   --wandb_entity=xxx \
   --wandb_watch all \
   --times_num 20 \
   --peft_type perCell_mag_add \
   --max_length 512 \
   --target_modules q_proj k_proj v_proj o_proj gate_proj up_proj down_proj

Using Pre-configured Scripts

We provide pre-configured training scripts for different tasks and the trainable parameters budget:

Commonsense Reasoning

# LLaMA-7B on commonsense tasks with top-20 parameters
bash scripts/percell/perCell_mag_add/LLaMA-7B/top20/commonsense.sh

# LLaMA-7B on commonsense tasks with top-1 parameters
bash scripts/percell/perCell_mag_add/LLaMA-7B/top1/commonsense.sh

Arithmetic Reasoning

# LLaMA-7B on commonsense tasks with top-20 parameters
bash scripts/percell/perCell_mag_add/LLaMA-7B/top20/math.sh

# LLaMA-7B on commonsense tasks with top-1 parameters
bash scripts/percell/perCell_mag_add/LLaMA-7B/top1/math.sh

Training Parameters

Parameter	Description	Options
-task	Task type	commonsense, math
-model	Base model path	yahma/llama-7b-hf, yahma/llama-13b-hf, meta-llama/Llama-2-7b-hf,meta-llama/Meta-Llama-3-8B
--peft_type	PEFT method	perCell_mag_add
--target_modules	Target modules for selecting parameters	See below
--times_num	Top-K input connection for each neuron	1, 5, 10, 20, etc.
-e	Number of epochs	-

Target Modules

Common target modules for different models:

• LLaMA/LLaMA2/LLaMA3: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Custom selection: You can specify any subset of these modules

Results

We provide Weights & Biases links to present our results reported in the paper. Below are the results on commonsense and arithmetic reasoning tasks.

🧠 Commonsense Reasoning Results

🏗️ Base Model	⚙️ Params (%)	🧩 BoolQ	💡 PIQA	🤔 SIQA	📖 HellaS.	🧍 WinoG.	🧮 ARC-e	🧠 ARC-c	📚 OBQA	🌟 Avg.
LLaMA (7B)	0.404%	73.1	85.4	80.9	94.3	84.3	87.8	71.7	84.2	82.7
LLaMA (7B)	0.020%	69.6	83.6	80.5	92.3	81.1	84.0	68.1	84.0	80.0
LLaMA (13B)	0.327%	73.3	87.9	82.7	96.0	86.9	90.2	77.1	88.6	85.3
LLaMA (13B)	0.016%	73.0	86.4	82.2	94.5	84.0	87.6	74.5	86.0	83.5
Llama2 (7B)	0.404%	73.6	86.5	81.1	94.8	87.8	89.1	75.9	85.6	84.3
Llama2 (7B)	0.020%	71.4	82.8	79.8	93.3	84.0	85.4	70.1	81.2	81.0
Llama3 (8B)	0.343%	75.0	89.3	83.0	96.5	89.2	93.0	82.9	89.6	87.3
Llama3 (8B)	0.017%	71.7	84.9	81.4	93.9	85.4	88.8	77.0	83.8	83.4

---

➗ Arithmetic Reasoning Results

🏗️ Base Model	⚙️ Params (%)	🔢 MulAri	📚 GSM8K	➕ AddSub	💧 AQuA	🧮 SinEq	📊 SVAMP	📘 MAWPS	🌟 Avg.
LLaMA (7B)	0.404%	96.0	36.5	92.4	22.0	94.1	53.2	84.5	68.4
LLaMA (7B)	0.020%	89.0	30.3	87.1	22.8	83.7	48.9	77.7	62.8
LLaMA (13B)	0.327%	97.5	43.9	92.2	21.7	93.9	61.4	89.1	71.4
LLaMA (13B)	0.016%	94.5	43.0	88.6	25.6	90.4	56.7	83.6	68.9
LLaMA2 (7B)	0.404%	97.8	39.8	91.9	20.5	96.3	54.2	89.5	70.0
LLaMA2 (7B)	0.020%	90.8	36.1	88.4	22.8	87.6	52.1	82.4	65.7
Llama3 (8B)	0.343%	99.7	47.8	92.7	27.6	95.7	60.4	88.7	73.2
Llama3 (8B)	0.017%	97.2	63.7	91.9	26.4	92.9	75.0	88.7	76.5

---

Citation

@inproceedings{zhang-etal-2025-neuroada,
    title = "{N}euro{A}da: Activating Each Neuron{'}s Potential for Parameter-Efficient Fine-Tuning",
    author = "Zhang, Zhi  and
      Shen, Yixian  and
      Cao, Congfeng  and
      Shutova, Ekaterina",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.555/",
    pages = "10960--10977",
    ISBN = "979-8-89176-332-6"
}

Acknowledgements

Our code is based on LLM-Adapters and LoReFT. We thank the authors for their valuable contributions to the open-source community.

License

This project is licensed under the MIT License - see the LICENSE file for details.