BoLT is a benchmark of surrogate models fitted to real LLM experiment data that enables evaluation of Bayesian and black-box optimization methods on multi-fidelity, multi-objective, high-dimensional LLM tasks.
hub
A Survey on Data Selection for Language Models
28 Pith papers cite this work, alongside 6 external citations. Polarity classification is still indexing.
28
Pith papers citing it
6
external citations · Pith
abstract
A major factor in the recent success of large language models is the use of enormous and ever-growing text datasets for unsupervised pre-training. However, naively training a model on all available data may not be optimal (or feasible), as the quality of available text data can vary. Filtering out data can also decrease the carbon footprint and financial costs of training models by reducing the amount of training required. Data selection methods aim to determine which candidate data points to include in the training dataset and how to appropriately sample from the selected data points. The promise of improved data selection methods has caused the volume of research in the area to rapidly expand. However, because deep learning is mostly driven by empirical evidence and experimentation on large-scale data is expensive, few organizations have the resources for extensive data selection research. Consequently, knowledge of effective data selection practices has become concentrated within a few organizations, many of which do not openly share their findings and methodologies. To narrow this gap in knowledge, we present a comprehensive review of existing literature on data selection methods and related research areas, providing a taxonomy of existing approaches. By describing the current landscape of research, this work aims to accelerate progress in data selection by establishing an entry point for new and established researchers. Additionally, throughout this review we draw attention to noticeable holes in the literature and conclude the paper by proposing promising avenues for future research.
hub tools
citation-role summary
background 1
other 1
citation-polarity summary
representative citing papers
HERMES provides a reusable hierarchical labeling substrate for pre-training data that reveals granularity-specific effects in data mixing rules during model training.
D³ introduces a dynamic directional graph-constrained framework that models sample interactions via loss dependencies to derive an optimized training sequence for LLMs.
Online SFT data selection acts as an implicit preference model, shifting refusal rates, verbosity, and sycophancy in directions predictable from the selected data's attribute mixture.
On-policy self-distillation with sampled demonstrations reduces rollout diversity by amplifying existing probability gaps in the base model, unlike ideal RL which preserves ratios among correct outputs.
Web graph centrality from Common Crawl supplies an orthogonal signal for pretraining data selection that improves language model performance when central and peripheral hosts are balanced.
Short-term data selectors in multi-stage LLM fine-tuning can slow future learning and increase forgetting, formalized as myopic selection with a proposed LHAS objective to address it.
High-Entropy Sum (HES) selects high-quality reasoning data for LLMs by summing entropy of the top highest-entropy tokens, matching full-dataset performance with top 20% in SFT and outperforming baselines in RFT and RL.
Controlled experiments show structured reasoning traces and higher-density math-domain samples improve mathematical reasoning more than pure executable code, with internal routing patterns reflecting these data effects.
SEED models data selection as Weighted Independent Set on a similarity graph, using node value calibration and local scale normalization to produce compact high-quality training subsets that outperform prior methods on instruction tuning and segmentation tasks.
Synthetic pre-pre-training on structured data improves LLM robustness to noisy pre-training, matching baseline loss with up to 49% fewer natural tokens for a 1B model.
CRAFT filters training data via source clustering and conditional target selection to bound KL divergence to validation distributions, yielding 43.34 BLEU on English-Hindi translation from 33M pairs while running over 40x faster than TSDS.
S2H-DPO generates hierarchical prompt-driven preference pairs to improve multi-image reasoning in VLMs while keeping single-image performance intact.
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training performance.
DUET is a global-to-local method that optimizes LLM training data mixtures via Bayesian optimization guided by influence-based selection and feedback from unseen evaluation tasks, with a regret bound showing convergence to the optimal mixture.
The study filters non-English Wikipedia, reveals quality problems, proposes a 4-level ranking, and shows filtered data matches or beats raw data in language modeling with largest gains for lower-quality editions.
DCLM-Baseline dataset lets a 7B model reach 64% 5-shot MMLU accuracy after 2.6T tokens, beating prior open-data models by 6.6 points on MMLU with 40% less compute.
StarCoder2-15B matches or beats CodeLlama-34B on code tasks despite being smaller, and StarCoder2-3B outperforms prior 15B models, with open weights and exact training data identifiers released.
HDS uses Soft Actor-Critic RL with a multi-objective reward (data quality, inter-domain loss influence, weight norms) for online data mixing in LLM pre-training, reaching target perplexity with 44% fewer iterations and 7.2% MMLU gain on The Pile.
An iterative bootstrapped self-filtering approach selects balanced clean and diverse subsets from noisy vision-language datasets to train improved CLIP models.
An audit finds language model filters and guardrails disproportionately suppress mentions of marginalized groups via lexical cues while failing to catch explicit harms.
Reinforcement learning teaches LLMs to assess their own capabilities more effectively than supervised fine-tuning, preserves original skills, generalizes out of distribution, and aids local-cloud routing and data selection.
SPARD defends LLMs from harmful fine-tuning attacks via alternating safety projections and relevance-diversity DPP data selection, reporting lowest attack success rates on GSM8K and OpenBookQA while keeping task accuracy.
Fine-grained metadata such as document quality indicators accelerate LLM pretraining when prepended, and metadata appending plus learnable meta-tokens recover additional speedup via auxiliary tasks and latent structure.
citing papers explorer
-
BoLT: A Benchmark to Democratize Black-box Optimization Research for Expensive LLM Tasks
BoLT is a benchmark of surrogate models fitted to real LLM experiment data that enables evaluation of Bayesian and black-box optimization methods on multi-fidelity, multi-objective, high-dimensional LLM tasks.
-
HERMES: A Multi-Granularity Labeling Substrate for Pre-training Data Mixtures
HERMES provides a reusable hierarchical labeling substrate for pre-training data that reveals granularity-specific effects in data mixing rules during model training.
-
D$^3$: Dynamic Directional Graph-Constrained Data Scheduling for LLM Training
D³ introduces a dynamic directional graph-constrained framework that models sample interactions via loss dependencies to derive an optimized training sequence for LLMs.
-
Online Data Selection Is Implicit Alignment
Online SFT data selection acts as an implicit preference model, shifting refusal rates, verbosity, and sycophancy in directions predictable from the selected data's attribute mixture.
-
On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
On-policy self-distillation with sampled demonstrations reduces rollout diversity by amplifying existing probability gaps in the base model, unlike ideal RL which preserves ratios among correct outputs.
-
Hubs or Fringes: Pretraining Data Selection via Web Graph Centrality
Web graph centrality from Common Crawl supplies an orthogonal signal for pretraining data selection that improves language model performance when central and peripheral hosts are balanced.
-
The Long-Term Effects of Data Selection in LLM Fine-Tuning
Short-term data selectors in multi-stage LLM fine-tuning can slow future learning and increase forgetting, formalized as myopic selection with a proposed LHAS objective to address it.
-
Unified Data Selection for LLM Reasoning
High-Entropy Sum (HES) selects high-quality reasoning data for LLMs by summing entropy of the top highest-entropy tokens, matching full-dataset performance with top 20% in SFT and outperforming baselines in RFT and RL.
-
What Really Improves Mathematical Reasoning: Structured Reasoning Signals Beyond Pure Code
Controlled experiments show structured reasoning traces and higher-density math-domain samples improve mathematical reasoning more than pure executable code, with internal routing patterns reflecting these data effects.
-
SEED: Targeted Data Selection by Weighted Independent Set
SEED models data selection as Weighted Independent Set on a similarity graph, using node value calibration and local scale normalization to produce compact high-quality training subsets that outperform prior methods on instruction tuning and segmentation tasks.
-
Synthetic Pre-Pre-Training Improves Language Model Robustness to Noisy Pre-Training Data
Synthetic pre-pre-training on structured data improves LLM robustness to noisy pre-training, matching baseline loss with up to 49% fewer natural tokens for a 1B model.
-
CRAFT: Clustered Regression for Adaptive Filtering of Training data
CRAFT filters training data via source clustering and conditional target selection to bound KL divergence to validation distributions, yielding 43.34 BLEU on English-Hindi translation from 33M pairs while running over 40x faster than TSDS.
-
S2H-DPO: Hardness-Aware Preference Optimization for Vision-Language Models
S2H-DPO generates hierarchical prompt-driven preference pairs to improve multi-image reasoning in VLMs while keeping single-image performance intact.
-
RewardBench 2: Advancing Reward Model Evaluation
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training performance.
-
DUET: Optimizing Training Data Mixtures via Feedback from Unseen Evaluation Tasks
DUET is a global-to-local method that optimizes LLM training data mixtures via Bayesian optimization guided by influence-based selection and feedback from unseen evaluation tasks, with a regret bound showing convergence to the optimal mixture.
-
How Good is Your Wikipedia? Auditing Data Quality for Low-resource and Multilingual NLP
The study filters non-English Wikipedia, reveals quality problems, proposes a 4-level ranking, and shows filtered data matches or beats raw data in language modeling with largest gains for lower-quality editions.
-
DataComp-LM: In search of the next generation of training sets for language models
DCLM-Baseline dataset lets a 7B model reach 64% 5-shot MMLU accuracy after 2.6T tokens, beating prior open-data models by 6.6 points on MMLU with 40% less compute.
-
StarCoder 2 and The Stack v2: The Next Generation
StarCoder2-15B matches or beats CodeLlama-34B on code tasks despite being smaller, and StarCoder2-3B outperforms prior 15B models, with open weights and exact training data identifiers released.
-
Holistic Data Scheduler for LLM Pre-training via Multi-Objective Reinforcement Learning
HDS uses Soft Actor-Critic RL with a multi-objective reward (data quality, inter-domain loss influence, weight norms) for online data mixing in LLM pre-training, reaching target perplexity with 44% fewer iterations and 7.2% MMLU gain on The Pile.
-
Data Selection Through Iterative Self-Filtering for Vision-Language Settings
An iterative bootstrapped self-filtering approach selects balanced clean and diverse subsets from noisy vision-language datasets to train improved CLIP models.
-
Epistemic Injustice in Language Models: An Audit of Pretraining Filters and Guardrails
An audit finds language model filters and guardrails disproportionately suppress mentions of marginalized groups via lexical cues while failing to catch explicit harms.
-
Capability Self-Assessment: Teaching LLMs to Know Their Limits
Reinforcement learning teaches LLMs to assess their own capabilities more effectively than supervised fine-tuning, preserves original skills, generalizes out of distribution, and aids local-cloud routing and data selection.
-
SPARD: Defending Harmful Fine-Tuning Attack via Safety Projection with Relevance-Diversity Data Selection
SPARD defends LLMs from harmful fine-tuning attacks via alternating safety projections and relevance-diversity DPP data selection, reporting lowest attack success rates on GSM8K and OpenBookQA while keeping task accuracy.
-
Beyond URLs: Metadata Diversity and Position for Efficient LLM Pretraining
Fine-grained metadata such as document quality indicators accelerate LLM pretraining when prepended, and metadata appending plus learnable meta-tokens recover additional speedup via auxiliary tasks and latent structure.
-
SEDD: Scalable and Efficient Dataset Deduplication with GPUs
SEDD delivers a distributed GPU deduplication system that reports up to 158x speedup over CPU baselines and 7.8x over NeMo Curator on 30M documents while preserving MinHash fidelity above 0.95 Jaccard.
-
Factual Inconsistencies in Multilingual Wikipedia Tables
The study introduces a method for detecting and categorizing cross-lingual factual inconsistencies in Wikipedia tables using alignment techniques and metrics on sample data.
-
From Cool Demos to Production-Ready FMware: Core Challenges and a Technology Roadmap
A semi-structured thematic synthesis identifies core challenges in FM selection, alignment, prompting, orchestration, testing, deployment, and cross-cutting concerns like observability for production-ready FMware.
-
Position: Multimodal Large Language Models Can Significantly Advance Scientific Reasoning
Position paper claims multimodal LLMs can significantly advance scientific reasoning and proposes a four-stage roadmap plus challenges and suggestions.