Target-Oriented Pretraining Data Selection
via Neuron-Activated Graph

Zijun Wang1,2, Haoqin Tu2, Weidong Zhou1, Yiyang Zhou3, Xiaohuan Zhou1,
Bingni Zhang1, Weiguo Feng1, Taifeng Wang1, Cihang Xie2, Fengze Liu1
1ByteDance, 2UC Santa Cruz, 3UNC-Chapel Hill
NAG teaser

General quality-based data selection is often misaligned with specific downstream capabilities (left), while prior target-oriented methods rely on shallow similarity to target examples (middle left). Our NAG instead aligns pretraining data with target tasks by selecting inputs that activate similar neurons in the LLM, capturing the underlying capability required for the target (middle right), even across different domains (e.g., economics vs. math).

+4.9%
avg. gain over Random
+5.3%
over SOTA on HellaSwag
0.12%
neurons → 23.5% drop
Training-Free
interpretable signal

Abstract

Everyday tasks come with a target, and pretraining models around this target is what turns them into experts. In this paper, we study target-oriented language model (LM) pretraining by introducing Neuron-Activated Graph Ranking (NAG-based Ranking), a training-free and interpretable framework for target pretraining data selection. Rather than using black-box representations, our approach directly characterizes each target input by a sparse set of high-impact neurons in any off-the-shelf LLMs.

Concretely, we quantify neuron impact and select the most influential neurons across layers into a compact Neuron-Activated Graph (NAG), and rank candidate data by NAG similarity to target examples. We conduct experiments across six benchmarks, where our NAG-based Ranking improves target-oriented pretraining by 4.9% on average over random sampling, and also outperforms state-of-the-art baselines by 5.3% accuracy on HellaSwag. It also remains effective under a more applicable multi-target setting, where our best setup surpasses two baselines by 1.1% and 4.1%, respectively.

Furthermore, we provide a comprehensive analysis on why and how our NAG works. For example, deactivating NAG-selected neurons (only 0.12% of all) causes a 23.5% performance collapse, and restricting NAG to the final layer incurs a 4.1% average drop, indicating that NAG captures a sparse "functional backbone" for learning target features.

Pipeline

NAG pipeline

Given a small set of target examples, we run them through a frozen off-the-shelf LLM, score the impact of every neuron, and keep the top-K per layer to form a compact NAG for each input. The target NAGs are aggregated into a per-layer neuron-activation profile. Every candidate document in the pretraining pool gets its own NAG and is ranked by its similarity to the target profile; the top-rf fraction is then selected for LLM pretraining.

Key Idea: Inputs that engage similar neurons in an LLM share similar task-relevant capabilities — so neuron co-activation is a direct, interpretable signal for picking pretraining data that matches a target.

Main Results

NAG-based Ranking consistently improves target performance across six benchmarks

Method ARC-C HellaSwag TriviaQA MMLU XStoryCloze XWinograd Avg.
Random 28.5 51.6 15.6 30.2 67.1 76.5 44.9
FineWeb-Edu 34.3+5.8 55.3+3.7 20.1+4.5 32.8+2.6 65.9−1.2 76.2−0.3 47.4+2.5
Single-Target
BETR 32.3+3.8 57.5+5.9 20.2+4.6 31.1+0.9 71.0+3.9 80.7+4.2 48.8+3.9
NAGQwen3-1.7B 34.0+5.5 60.6+9.0 22.3+6.7 32.2+2.0 70.0+2.9 80.1+3.6 49.8+4.9
NAGLlama-3.2-3B 35.0+6.5 58.6+7.0 21.3+5.7 31.5+1.3 70.8+3.7 80.6+4.1 49.6+4.7
NAGSmolLM3-3B 35.0+6.5 59.8+8.2 22.6+7.0 31.2+1.0 70.5+3.4 80.6+4.1 49.9+5.0
Multi-Target
BETR 30.3+1.8 49.3−2.3 11.6−4.0 29.9−0.3 69.5+2.4 76.1−0.4 44.4−0.5
NAGQwen3-1.7B 33.4+4.9 57.8+6.2 19.2+3.6 31.5+1.3 69.3+2.2 79.9+3.4 48.5+3.6
NAGLlama-3.2-3B 32.0+3.5 54.9+3.3 18.0+2.4 31.4+1.2 69.8+2.7 79.9+3.4 47.6+2.7
NAGSmolLM3-3B 31.8+3.3 55.2+3.6 19.9+4.3 30.6+0.4 69.2+2.1 80.2+3.7 47.8+2.9

Bold = best overall, underline = second overall, shade = best within Single-/Multi-Target setting. Subscripts are deltas vs. Random (red for gains, blue for drops).

Observation 1: NAG-based Ranking yields a +4.9% average gain over Random and beats both quality-focused (FineWeb-Edu, +2.4%) and target-oriented (BETR, +1.0%) baselines, with +5.3% on HellaSwag.

Observation 2: The gain is robust across backbone models (+4.7% to +5.0%) and remains effective under the more challenging Multi-Target setting, where BETR drops -4.4% but NAG still gains +3.6%.

NAG Complements Existing Quality Signals

Method ARC-C HellaSwag TriviaQA MMLU XStoryCloze XWinograd Avg.
FineWeb-Edu 34.3 55.3 20.1 32.8 65.9 76.2 47.4
+ NAGQwen3-1.7B 35.3+1.0 57.7+2.4 21.7+1.6 32.5−0.3 67.2+1.3 79.2+3.0 48.9+1.5
+ NAGLlama-3.2-3B 35.2+0.9 57.4+2.1 21.7+1.6 32.7−0.1 68.2+2.3 78.6+2.4 49.0+1.6
+ NAGSmolLM3-3B 35.7+1.4 58.1+2.8 22.7+2.6 33.1+0.3 69.0+3.1 78.9+2.7 49.6+2.2

shade = best per column. Subscripts are deltas vs. FineWeb-Edu.

Observation 3: Adding NAG signals on top of FineWeb-Edu further boosts average accuracy by +1.8%, indicating NAG captures information complementary to general-quality scores.

Why NAG Works

A Sparse Functional Backbone

Method ARC-C HellaSwag TriviaQA MMLU XStoryCloze XWinograd Avg.
Qwen3-1.7B-Base 55.7 66.9 36.3 45.9 72.4 86.5 60.6
Deactivate 20 neurons per layer (0.12%)
Deactivate Random 55.5−0.2 66.8−0.1 35.8−0.5 45.8−0.1 72.4−0.0 85.9−0.6 60.4−0.2
Deactivate NAG 30.4−25.3 45.6−21.3 0.3−36.0 29.1−16.8 56.9−15.5 60.6−25.9 37.1−23.5

Subscripts are performance drops relative to the original Qwen3-1.7B-Base (blue for drops, gray for negligible).

Observation 4: Deactivating just 0.12% NAG-selected neurons triggers a 23.5% performance collapse, while deactivating the same number of random neurons has near-zero effect — NAG isolates a sparse functional backbone in the LLM.

A Task-Discriminative Representation

t-SNE of NAG

t-SNE visualization of NAG-based representations across ten datasets. Clusters align tightly with task identity, and their relative positions reflect task relevance (e.g., MathQA and GSM8K cluster near each other yet far from XNLI).

Observation 5: NAG features form clean task-level clusters, evidencing that they encode task-discriminative structure rather than surface lexical similarity.

A Utility-Aligned Ranking

Filter rate

Performance under varying filtering rates rf. As rf decreases (only the highest-ranked data is kept), Random and BETR degrade, while NAG improves, peaking at rf = 5% with a +1.8% additional gain.

Observation 6: NAG's induced ranking is tightly aligned with downstream task utility — aggressive filtering of low-ranked samples actually improves accuracy.

How NAG Operates

Best with FFN UP-projection Neurons

Neuron type

up_proj neurons (60.6%) clearly beat residual-interface projections like down_proj (58.0%) and k_proj (56.7%), suggesting expansion layers in higher-dimensional latent space best isolate task-specific signals.

Observation 7: NAG works best when constructed from FFN up_proj neurons — the high-dimensional expansion latent space exposes task-discriminative activation patterns that compressed residual-stream projections (down_proj, k_proj) wash out.

Aggregating Signals Across All Layers Matters

Method ARC-C HellaSwag TriviaQA MMLU XStoryCloze XWinograd Avg.
NAGAll Layer 34.0 60.6 22.3 32.2 70.0 80.1 49.8
NAGLast Layer 30.5−3.5 55.2−5.4 15.1−7.2 29.9−2.3 67.8−2.2 75.5−4.6 45.7−4.1

Subscripts are deltas vs. NAGAll Layer (blue for drops).

Observation 8: Restricting NAG to the final layer incurs a -4.1% drop — task-relevant signals are distributed across all layers, not just the top.

A Highly Sparse Set of Neurons Is Enough

NAG width

Performance peaks at rk ≈ 0.3% across model scales (1.7B, 4B, 8B). Increasing sparsity 7× brings little extra gain, confirming the most competent task-relevant signals are concentrated in a very sparse set of high-impact neurons.

Observation 9: NAG concentrates into roughly 0.3% of neurons per layer, and larger LLMs yield more discriminative neurons under the same sparsity.

BibTeX

@misc{wang2026targetorientedpretrainingdataselection,
      title={Target-Oriented Pretraining Data Selection via Neuron-Activated Graph},
      author={Zijun Wang and Haoqin Tu and Weidong Zhou and Yiyang Zhou and Xiaohuan Zhou and Bingni Zhang and Weiguo Feng and Taifeng Wang and Cihang Xie and Fengze Liu},
      year={2026},
      eprint={2604.15706},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.15706},
}