Organizations: Department of Computer Science University of Chicago · Department of Linguistics The University of Texas at Austin · Allen Institute for Artificial Intelligence
Selecting high-quality training data can reduce computation cost for LLM fine-tuning. Prior work on data selection developed a variety of difficulty-based scores, motivated by the intuition that harder examples are more informative. While effective for smaller classification models, we find that such methods often fail to outperform random selection for LLM fine-tuning. Our analysis reveals that random selection performs better because it is more representative of the underlying input space---i.e., has better coverage. To validate the importance of coverage, we implement a simple clustering-based selection method which selects data that is more representative of the underlying input distribution. On a set of diverse single-domain tasks, this method yields performance at or above that of random selection. On the diverse, multi-domain Tülu 3 mixture, where random sampling no longer guarantees a balanced subset, our method outperforms all methods. Overall, our results indicate that prioritizing coverage over difficulty leads to more effective data selection for LLM fine-tuning.
Figures & tables
Figure 1: Comparative performance of difficulty-based data selection methods on the Llama 3 8B and OLMo 2 7B models across the three datasets, as a function of the percentage of training data selected. For random selection, the points show the mean test accuracy and shaded regions indicate the ±1 standard deviation band across runs. The blue dashed line indicates results from fine-tuning on 100% training data, while the gray dotted line indicates zero-shot performance. We repeat all our experiments with difficulty-based metrics for 3 random runs where the randomness mainly comes from the training.
Figure 2
Figure 3: Performance of the Coverage-based selection method on Llama 3 8B and OLMo 2 7B models as a function of training data selected, along with comparisons to random selection, the best overall difficulty-based method (VoG), and D2 . For random selection and Coverage, the points show the mean test accuracy and shaded regions indicate the ±1 standard deviation band across runs. The blue dashed line indicates results from fine-tuning on 100% training data, while the gray dotted line indicates zero-shot performance.
MMLU (Full: 62.2)
BBH (Full: 48.5)
HumanEval (Full: 45.9)
Method
1%
5%
10%
15%
25%
1%
5%
10%
15%
25%
1%
5%
10%
15%
25%
VoG
61.0
59.6
60.9
61.1
61.2
45.7
46.8
45.5
47.4
48.3
30.5
33.6
32.3
30.5
32.3
Variability
61.6
61.8
61.8
61.6
61.9
0 7.5
49.7
49.4
48.8
47.9
14.9
31.5
31.9
30.9
30.6
Perplexity
61.6
61.0
61.6
61.5
61.7
11.5
47.2
48.7
49.0
48.3
0 6.5
30.3
33.1
32.6
31.4
Confidence
61.7
61.6
61.5
61.3
62.1
45.2
49.5
49.1
48.0
48.0
15.7
27.3
31.3
29.9
29.5
Random
60.9
61.2
61.6
61.9
61.7
47.4
48.8
48.5
49.6
48.2
30.8
36.5
31.1
36.4
37.5
Table 2: Downstream accuracy (%) on OLMo 2 7B fine-tuned on subsets of Tülu 3, across selection percentages. Bold marks the best method per column; underline marks any method that meets or exceeds the full fine-tuning baseline (in header).
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Effect of random sampling vs. selection using perplexity after performing clustering for our coverage-based method.
Dataset
Model
Zero-shot
100%
Social IQa
Llama 3 8B
0.633
0.755
OLMo 2 7B
0.665
0.782
CommonsenseQA
Llama 3 8B
0.667
0.784
OLMo 2 7B
0.678
0.812
CosmosQA
Llama 3
0.606
0.853
OLMo 2 7B
0.747
0.884
Appendix
Table 3: Datasets considered for experiments. Zero-shot performance and performance on using the entire train data on Llama 3 and OLMo 2
Dataset
Model
Total tokens
Social IQa
Llama 3 8B
1,786,531
OLMo 2 7B
1,786,578
CommonsenseQA
Llama 3 8B
577,187
OLMo 2 7B
577,230
CosmosQA
Llama 3 8B
3,692,490
OLMo 2 7B
3,692,921
Appendix
Table 4: Total tokens.
Figure 5: Data Selection FLOPs ( ×1015 ) vs. Average Accuracy (calculated over all percentages) across all methods. X axis is in logarithmic scale. Dashed line indicates Random Selection performance which has a data-selection FLOP value of 0.
Difficulty Score
Dataset
Data Selection FLOPs ( ×1015 )
Perplexity
Social IQa
25
CommonsenseQA
8.08
CosmosQA
51.7
Confidence
Social IQa
225
CommonsenseQA
72.7
CosmosQA
465
Appendix
Table 5: FLOPS utilized for data selection using Llama 3 8B.
Difficulty Score
Dataset
Data Selection FLOPs ( ×1015 )
Perplexity
Social IQa
28.6
CommonsenseQA
9.23
CosmosQA
59.1
Confidence
Social IQa
257
CommonsenseQA
83.1
CosmosQA
532
Appendix
Table 6: FLOPS utilized for data selection using OLMo 2 7B.
Figure 6: Per-dataset and average accuracy across different values of the clustering parameter k at 25% selection.
Figure 7: Comparative performance of 10% selected data on the Llama 3 8B and OLMo 2 7B models across all the datasets. Accuracy on CommonsenseQA, CosmosQA, and Social IQA using Last Layer vs. Middle Layer embeddings in the clustering step of our coverage based method.
Data selection chooses informative, representative, or reliable examples to improve model training and evaluation while reducing data, compute, or labeling costs.