- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper focuses on the targeted instruction data selection problem. It first shows that current representation-based selection methods suffer from redundancy issues, then proposes two methods: CRDS-R and CRDS-W to alleviate these issues. CRDS-R applies Rademacher projection to multiple late hidden layers and concatenates them to preserve the original dimensionality. CRDS-W uses whitening-based dimensionality reduction and plugs the resulting representations into the selection pipeline. Experiments on two models across five benchmarks show the effectiveness of CRDS-R and CRDS-W.
Critical strengths
-
The paper is easy to follow and understand. The technical part is straightforward. Using embeddings is much more efficient than computing gradients. The experiments are thorough, and the ablation studies provide good support for the main claims.
-
The paper conveys an important message: selection performance depends not only on the selection strategy but also on the embedding space.
Critical weaknesses
-
To support the redundancy claim, a single setting on GSM8K is clearly not enough. The authors should at least evaluate this across multiple benchmarks. Furthermore, the connection between reducing redundancy and improving selection performance is not well established. For example, CRDS-R concatenates multiple layers’ embeddings, so the performance gain could simply come from incorporating more information. A similar concern applies to CRDS-W: the gain could be explained by a better distance metric rather than by reduced representation redundancy. The authors are encouraged to clearly define what “redundancy” means and explain how exactly CRDS-R and CRDS-W reduce it.
-
The compared baseline methods are limited. Some gradient-based methods are missing, such as LESS [R1] and TSDS [R2].
-
Rademacher projection, multi-layer embedding concatenation, and whitening-based dimensionality reduction are all existing techniques, which makes the technical novelty seem limited. The main contribution of the paper appears to be verifying the effectiveness of these techniques for large-scale targeted instruction data selection.
[R1] LESS: Selecting Influential Data for Targeted Instruction Tuning. ICML 2024
[R2] TSDS: Data Selection for Task-Specific Model Finetuning. NeurIPS 2024
Questions for authors
-
In practice, how should one decide which method to use, CRDS-R or CRDS-W? Is there any measurable property of the representations that predicts whether CRDS-R or CRDS-W will help?
-
In Table 3 (Qwen-2.5-7B setting), CRDS-W achieves 87.57 on GSM8K, but in Table 25 the highest GSM8K number is 86.58. Are the configurations for these two tables different?
-
How sensitive are the results to the size and composition of the query set?
3 comments
Thanks for clarifying why you ran the binarization experiment. However, I still think GSM8K alone is too little evidence for a broad claim about redundancy in LLM embeddings. CRDS does improve selection on other tasks, but I’m less convinced about why it works. CRDS-R can draw on information from several layers, and whitening in CRDS-W changes which examples are considered similar. Either could help explain the gains. The rebuttal doesn’t add an experiment that would let me tell how much of the improvement comes from reducing redundancy, so I’m keeping my score at 2.
We thank the reviewer for the response and would like to provide some further clarification regarding our paper.
Binarization experiments
We agree with your point. In the current version, the discussion of binarization is somewhat too detailed, which may make it appear to be a major component of our empirical evaluation. In fact, we do not intend to emphasize that binarization makes a substantial contribution to improving the experimental results; intuitively, the method is quite simple, and our preliminary experiments did not show a large improvement.
Instead, we view binarization as a useful starting point, and we will reduce the space devoted to it in the paper. Before this work, RDS+ was the strongest semantic embedding method we were aware of, and it used a large language model encoder rather than a dedicated semantic model such as Sentence-BERT. The fact that such a simple binarization method could improve semantic representations extracted by a large language model encoder suggested to us that these encoders have potential worth exploring. This motivated us to conduct a series of further investigations.
Mechanism of CRDS
We acknowledge that the paper does not provide a sufficient analysis of the mechanism behind CRDS, and we consider this a limitation of our work. However, Section 5.2 does discuss the possible mechanism of CRDS-W.
We believe that whitening may act as a form of regularization that removes low-variance components and focuses similarity computation on features with more pronounced variation in the data. The singular-value spectrum after binarization is similar to that after whitening, and binarization removes information about vector magnitudes. This suggests that the two methods may have similar regularization mechanisms.
For CRDS-W, we acknowledge that its effect may fundamentally arise from a change in the metric, since the coordinate system has changed. However, CRDS-W uses substantially fewer encoded dimensions than the original embeddings, and dimensionality reduction discards many dimensions. From this perspective, we believe it is reasonable to argue that the embeddings contain redundancy.
For CRDS-R, we acknowledge that both feature fusion and feature projection may contribute to the improvement. We agree that additional ablation experiments would help demonstrate this trend. Unfortunately, a change in the primary authors’ institutional affiliations has left us without the budget for additional experiments. These experiments require substantial GPU resources, and we are currently unable to fund the necessary compute.
We thank the reviewer for the detailed comments. Below, we clarify the role of our preliminary experiments, the constraints on baseline comparisons, and the contributions of our work.
W1. Evidence of representation redundancy
The binarization experiments were intended as an initial observation motivating this work, rather than direct proof of representation redundancy. The finding that binarized features could improve SFT data selection led us to investigate redundancy in textual representations. The subsequent CRDS experiments provide further empirical support for this hypothesis.
We did not allocate additional experimental resources to binarization because it is a relatively simple approach and the observed improvements were modest. Instead, we focused our resources on developing and evaluating CRDS.
W2. Comparisons with LESS and TSDS
We acknowledge that comparisons with LESS and TSDS are missing. Our work follows the large-scale setting of RDS+, whose authors identified gradient extraction as prohibitively expensive for datasets containing millions of examples. After assessing the computational requirements of these comparisons, we concluded that they exceeded our available resources at the scale studied in this paper.
Nevertheless, our experiments demonstrate clear improvements over RDS+, a strong and practically relevant baseline: it is the SFT data-selection method used in the official Tulu 3 release. These results establish an improvement over a method already used in large-scale model development, although they do not substitute for direct comparisons with LESS and TSDS.
W3. Novelty
We acknowledge that both algorithms in CRDS draw inspiration from existing methods. Our motivation for adopting these techniques is to address redundancy in representations produced by large-model encoders. Our contributions lie in the following findings:
Compared with pretraining datasets, which are orders of magnitude larger, SFT datasets are small enough to accommodate algorithms with higher computational complexity. We believe that more refined algorithmic designs could further improve SFT data selection and potentially reveal a scaling law for this task.
Q1. Applicability of the methods
As shown in our experiments, CRDS-W performs better on the 16B model, whereas CRDS-R performs better on the 8B model. At present, we do not have a reliable criterion for predicting which method will perform better in a new setting. We therefore recommend evaluating both methods, when resources permit, and choosing the one that performs better for the intended task.
Q2. Experimental settings
We checked the original experimental records and confirmed that the discrepancy reflects randomness in the evaluation results rather than a difference in experimental settings. This configuration was evaluated twice: once during the initial exploration and once during the subsequent batch of ablation experiments. The main table reports the earlier result.
Based on our records, this discrepancy is confined to this data point. The other configurations were not evaluated in separate repeated runs because of our limited computational budget.
Q3. The query set
We have not systematically investigated how the characteristics of the query set affect performance. Empirically, our methods tend to perform well with smaller query sets that have distinctive task characteristics, such as GSM8K. For larger query sets, such as MMLU, the improvements are less consistent. Overall, we believe this is an important question that deserves further attention, and we will continue to investigate it in our work.