acceptodds
Withdrawn from ICLR 2027

Beyond Group Proportions: Quality Filtering and Dialect Gaps

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Quality filters decide which documents survive into a language model's training data, and their effects fall unevenly across varieties of English. The usual way to check for this is to count how much data each group keeps. We ask whether that count tells the right story for African American English (AAE), the variety spoken by many Black Americans, by comparing models trained on datasets that contain exactly the same number of AAE-aligned and white-aligned documents but different documents within each group. In a controlled continued-pretraining experiment on tweets, replacing the documents chosen by the DataComp-LM quality classifier with different documents at the same group counts removes 63% of the widening in the likelihood gap between the groups relative to uniform selection, and it improves fit to both groups relative to the filter's own documents. Restoring the original group share also narrows the gap, so the count matters, but it is not the whole account. In web pretraining from scratch at a matched token budget, an additional cut with the same classifier on an already-curated corpus raises the between-group perplexity ratio by 2.9%, whereas a second pool and budget give an estimate whose interval includes zero. Reweighting documents kept by a perplexity filter does not reproduce the benefit of changing them. These results argue for auditing which examples of a variety remain in the training data, and not only how many.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
language models, pretraining data, quality filtering, data selection, dialect fairness, African American English

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_eVCQOpenReview ↗
Rating
1
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper makes an important, interesting, and well-supported point that the ratio alone between dialects is not the only factor in performance discrepancies. In other words, even at the same data ratio, data selection can result in massive difference in the performance gaps between AAE and white English. They perform pre-training experiments on GPT 2 using a dialect-tagged twitter corpus as well as we corpora. Their findings are significant when it comes to convincing those who are pre-training to assess data makeup beyond frequency, especially because for many dialects frequency is not something that is easy to fix. I think with a complete re-write to make it more clear, but with largely the same experiments, this would make a good conference paper but at this stage it is too confusing.

Critical strengths

The experiments in this paper are very well done - many seeds, lots of baselines, and diverse data sources. I believe that the findings are also novel and interesting. While their results are discussing just AAE, it is easy to see that they would likely transfer to other dialects/varieties. They primarily focus on the NLL gap between the dialects, which concerned me at first that part of the reduced gap was due to worse white English performance, but it seems that they address that in the main body and for the most part it is not a concern.

Critical weaknesses
  1. Honestly, my biggest gripe is how difficult this paper was for me to parse. I found myself having to jump around a lot to figure out what was being referred to and relying heavily on figures/tables to decode the prose. The issue starts with the abstract, which I found to be very dense with numeric results shown without sufficient background to understand them. Through the introduction and results, there is no unified list of all the data variations with their differences, leading me to need to put it together with a mix of close reading and making assumptions based on the names of the methods. I'm still not sure what ClusterRandom is and whether this is a new method proposed by the authors. As someone not familiar with pre-training data filtering, I had no background on DataComp-LM and the paper provided no insight, forcing me to go read that paper to understand what the filtering process was. This issue continues on throughout the paper and means I can't really verify the claims.

  2. I can't tell if this is an argument against filtering in general, filtering in a manner that restricts AAE more heavily due to bias, or against combining filtering and uneven data distributions. For example, uniform, unfiltered seems to perform the best suggesting that we shouldn't filter and instead just use all the data. While the paper seems to make an argument for not filtering, the restored-share baseline seems to suggest that the filtering isn't as much of a problem as the ratio, with fixed-share performing worse. This is especially true because it isn't stated how much the filter cut non-AAE data.

  3. The web comparison section seems to not have any explicit information about dialect - even from a dialect classification system. As a result, it's hard to connect these results to the broader argument from the previous section. I believe your argument would be stronger if you explicitly were able to point towards specific attested AAE constructions that are filtered out and demonstrate that the performance drop is due to that. Otherwise its hard to say whether this is due to the filtering causing a ratio discrepancy (the ratio-share comparison) or the filtering removing specific features that are in the evaluation set (the fixed-share comparison).

Questions for authors
  1. For Table 2, it seems that you have only sampled one dataset for the comparison, but you ran multiple seeds (with different data sampling?), so do these differences hold when looking at the standard deviation of the values over sampling runs?
  2. What is the motivation behind ClusterRandom? You are semantically embedding these documents and proportionally sampling from them, which introduces a whole different axis. I would understand more if you ensured each cluster contributed uniform dialectal proportions, to basically control for topic and dialect when creating the dataset, but that doesn't seem to be the case. Also you need to introduce this earlier if it isn't a pre-established method.
Reviewer_S2NZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the effects that quality filtering can have on the representation of different groups, both by changing the distribution of text and by which kinds of texts are chosen. The authors experiment with AAE and White-American-aligned data on two experimental setups, and show that the generally used data selection setup leads to a gap between the two groups. Their most important finding is that reporting representation as percentages does not adequately reflect the documents that were chosen and the resulting model.

Critical strengths

Topic and methods: Representativeness of data after quality filtering is of utmost importance, specifically now that new training data is sparse and synthetic/LLM-rewritten data is on the rise. We cannot have documents deemed most informative that are aligned with only one or a handful of groups. The methodology introduced here can be applied to other groups as well, as long as a labelled dataset is available.

Critical weaknesses

Clarity: Soundness score of 2 due to the fact that it is very hard to understand certain results, although the main takeaways can be deciphered. Many terms/methods are introduced, then skipped over, and definitions are given a few pages later (e.g., ClusterRandom (127-> vs. Sec.4; line 049 vs. Equation 1) with no reference. Certain results are omitted(?) from figures, which makes them difficult to understand. These things should be taken care of, especially since the paper is under the maximum allowed length, so the lack of space is not a problem.

Questions for authors
  1. What is the “existing lexical proxy” (line 155 and Table 2 caption)?
  2. What is the motivation for ClusterRandom, and how does it relate to the other selection criteria? Lines 127 to 131 feel totally disjoint from the surrounding text, and Sec. 4 mostly describes the implementation.
  3. What does “region-matched” (e.g., line 208) mean concretely?
  4. (as you have space) Can you add figures that include all selection criteria and the effect of rewriting?
Reviewer_NbyZOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper explores the biases of the DCLM quality classifier on the African American English dialect, comparing it with white-aligned documents. Additionally, authors include experiments using the GPT-2 perplexity scorer. Authors perform the analyses on 124M, 150M, and 1B models by measuring the mean negative log likelihood gap per token, and for 124M and 150M models on web-text data the perplexity on WikiText-103, negative log likelihood on C4, and accuracy on LAMBADA. The main finding is that changing the retained documents can improve the model fit measured by the negative log likelihood gap even when group counts remain fixed.

Critical strengths
  1. The problem of ensuring fairness across dialects of a language is well-motivated and its findings could potentially also be useful even further for ensuring fairness between higher-resource and lower-resource languages in large language model training.
  2. Author compare several interventions: fixed-share, restored-share, ClusterRandom.
  3. The paper compares both the fairness gap via the negative log likelihood, as well as the quality on WikiText-103, C4, and LAMBADA, and runs training over multiple seeds with confidence intervals.
Critical weaknesses
  1. There is no group-stratified uniform control experiment and without it, the fixed-share result may mean that random selection within the group beats selecting top-scored documents, and although the authors note it, it takes away from the main result of the paper.
  2. Fixed-share training is never run on web-data pretraining or at larger scale, making the main result of the paper based on the first experiment consisting of tweets at maximum sequence length 64 and potentially not generalizing to web data.
  3. Findings would be stronger if the analysis was run on other dialects of different languages, or at least for other dialects of English.
  4. The readability of the paper could be improved. The paper is dense and the setups are described across many sections and even appendices. For example, Section 2 uses ClusterRandom, but the full definition is in Section 4. Similarly, many other details are in appendices, while the paper still has one unused page before hitting the limit of 9 pages. This does not take away from the results of the paper but rewriting it would help the reader understand the setup and its results better.
Questions for authors

I do not have any questions for the authors.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.