Research map / Large language models / Instruction and fine-tuning
Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoNA, which selects instruction-tuning data by comparing sparse autoencoder-disentangled (monosemantic) neuron activation embeddings of source and target samples using a generalized Jaccard similarity.
Problem
Selecting the most task-relevant subset of a general instruction dataset for fine-tuning LLMs to perform well on specific target tasks, given only a few target examples; existing influence-based methods are unstable and distribution-alignment methods rely on data-centric representations that ignore
Approach
Represent each sample by its neuron activation pattern in a pretrained LLM, disentangle polysemantic neurons with a TopK sparse autoencoder into monosemantic features, mean-pool token activations to remove length bias, and select samples by generalized Jaccard similarity to a task prototype.
Key results
- On LLaMA3.1-8B with 5% of OpenHermes-2.5, MoNA averages 55.75 vs 54.74 for the next-best baseline (DSIR) and 54.23 for full-data fine-tuning.
- On OLMo-7B with 5% of LESS, MoNA averages 34.71 vs 33.52 for LLM2Vec and 33.49 for LESS.
- On LLaMA2-13B with 5% of OpenHermes-2.5, MoNA averages 46.65 vs 46.18 for BM25.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select data with MoNA instead of the strongest reported baseline | LLaMA3.1-8B instruction tuning on OpenHermes-2.5 · OpenHermes-2.5, 5% selection | average score: 54.70 -> 55.75 | Helped |
| Select data with MoNA instead of the strongest reported baseline | LLaMA3.1-8B instruction tuning on Less · Less, 5% selection | average score: 66.08 -> 67.20 | Helped |
| Use SAE-derived monosemantic activations instead of polysemantic activations (ablation) | MoNA · Less | average score: 65.95 -> 67.20 | Helped |
| Add SAE mapping to raw hidden states (ablation) | MoNA · Less | average score: 65.19 -> 67.20 | Helped |
| Use Jaccard instead of Cosine similarity (ablation) | MoNA · OpenHermes-2.5 | average score: 63.78 -> 66.29 | Helped |
| Use Jaccard instead of Euclidean similarity (ablation) | MoNA · OpenHermes-2.5 | average score: 63.46 -> 66.29 | Helped |
| Increase sparsity parameter K from 24 to 192 (ablation) | MoNA · OpenHermes-2.5 | GSM8K: 64.29 -> 67.93 | Helped |
| Extract activations from layer 31 rather than layer 8 (ablation) | MoNA · OpenHermes-2.5 | average score: 54.39 -> 55.75 | Helped |
| Use mean aggregation instead of sum aggregation (ablation) | MoNA · OpenHermes-2.5 | GSM8K: 64.06 -> 67.93 | Helped |
| Select task-similar rather than dissimilar training data (ablation) | LLaMA3.1-8B instruction tuning · OpenHermes-2.5 | GSM8K: 33.06 -> 67.93 | Helped |
Surprising findings
- Selecting 10% of data performed worse than 5% in the reported experiments, which the authors attribute to dilution by less relevant samples.
- Sum aggregation of token activations induces length bias; mean aggregation (length normalization) was needed for good performance.
Limitations
- Only evaluated on text modality; multimodal extension not explored.
- Pretraining data selection not explored.
- Results vary in stability across backbones; task-level stability decreases on OLMo-7B and LLaMA2-13B.
Experimental setup
Datasets: OpenHermes-2.5 LESS MMLU GSM8K BBH MBPP GPQA TydiQA
Benchmarks: MMLU GSM8K BBH MBPP GPQA TydiQA
Metrics: accuracy exact match pass@1 F1 average score recall rate (Hit@10)
Baselines: BASE FULL RANDOM MATES LESS BM25 DSIR DLRDS-BGE DLRDS-LLaMA3-8B LLM2Vec
Builds on: LESS DSIR MATES CATS TopK sparse autoencoder
Closest papers by meaning
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.89)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.89)
- Neural Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data (NeurIPS 2025 · similarity 0.88)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.87)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.87)
- CoIDO: Efficient Data Selection for Visual Instruction Tuning via Coupled Importance-Diversity Optimization (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.