Skill-it! A Data-Driven Skills Framework for Understanding and Training Language Models

The quality of training data impacts the performance of pre-trained large language models (LMs). Given a fixed budget of tokens, we study how to best select data that leads to good downstream model performance across tasks. We develop a new framework based on a simple hypothesis: just as humans acquire interdependent skills in a deliberate order, language models also follow a natural order when learning a set of skills from their training data. If such an order exists, it can be utilized for improved understanding of LMs and for data-efficient training. Using this intuition, our framework formalizes the notion of a skill and of an ordered set of skills in terms of the associated data. First, using both synthetic and real data, we demonstrate that these ordered skill sets exist, and that their existence enables more advanced skills to be learned with less data when we train on their prerequisite skills. Second, using our proposed framework, we introduce an online data sampling algorithm, Skill-It, over mixtures of skills for both continual pre-training and fine-tuning regimes, where the objective is to efficiently learn multiple skills in the former and an individual skill in the latter. On the LEGO synthetic in the continual pre-training setting, Skill-It obtains 36.5 points higher accuracy than random sampling. On the Natural Instructions dataset in the fine-tuning setting, Skill-It reduces the validation loss on the target skill by 13.6% versus training on data associated with the target skill itself. We apply our skills framework on the recent RedPajama dataset to continually pre-train a 3B-parameter LM, achieving higher accuracy on the LM Evaluation Harness with 1B tokens than the baseline approach of sampling uniformly over data sources with 3B tokens.

Paper

References (75)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer XvRR7/10 · confidence 3/52023-06-20

Summary

This paper presents a way to introduce a curriculum of tokens to train LLMs. To do that, the paper introduces the concept of skills, i.e. behaviors that can be acquired by training the LLM in a subset of data, and ordered skill sets: partial orderings of skills that can be learned faster when using the skill preconditions. The paper shows that both in real and synthetic datasets, following the skill order allows for more efficient learning. Next, the paper presents techniques to sample data points from skills, given an ordered skill set. The apporach is evaluated on different datasets and model sizes, obtaining significant reductions in validation loss. Finally, the paper presents an approach to learn the skill ordering.

Strengths

- Relevance and novelty: The paper addresses an important problem; given the computational cost of training LLMs, optimizing performance given a fixed compute budge is an important problem. It also presents a new framing to do so, that departs from curriculum learning approaches and focuses in selecting data to learn a set of skills. The approach to sample skills is also novel and well thought, dealing with both sampling data that both improve the weaker skills and selects the amount of data to sample from each skill. - Method: Starting from a reasonable optimization objective and constraint (reduce the validation loss on every skill at every round), the method derives a closed from sampling rule, which is fast and simple to compute. The proposed rule can be applied to multiple cases: 1) those where all the training skills have to be evaluated, 2) cases where only a subset are considered for eval and 3) cases where no data of the eval skill is provided during training. - Generally (see questions) clear paper writing, the definitions and discussion on how different kinds of graphs affect good sampling strategies provides great intuition for the rest of the paper. The walkthrough examples in section 3.2 also clarifies the intuition of skill curriculum. - While the graph learning approach is computationally expensive, it can be applied to subsets of the dataset, allowing to transfer the task dependency matrix to the full data for sampling. - Experimental results: The method is tested across different datasets and model sizes. The proposed approach improves upon different sampling baselines, including methods that sample uniformly across skills, by obtaining lower validation losses or comparable losses than the baselines in a shorter amount of time. The method also improves over baseliens on out of domain settings, where the test skills are not available during training.

Weaknesses

- The method relies very strongly on having a task dependency graph. While the paper proposes two approaches to obtain these graphs, I would like further discussion into that, particularly on how to obtain the edge weights, and the robustness in the appraoch to errors or noise in this depnedncy graph. Clarity on writing: - Definitions: In the definition of skill it would be good to clarify what the improved metric is with respect to. Is it compared to training from a random subset of X? - Similarly, it only becomes clear in Section 4.1 that the goal is to learn a subset of skills (which can include only 1 skill or the full skill training set). It would be good to clarify this better in the introduction.

Questions

- On the definition skill, it would be good to clarify what the improved metric is with respect to. Is it compared to training from a random subset of X? - I understand from 4.4 how to learn the presence or absence of edges in the graph, but how do we find from there the weight of that edge? How sensible is the present algorithm to the edge weight.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

No limitation is mentioned in the paper.

Reviewer 6a648/10 · confidence 5/52023-06-30

Summary

This paper addresses the problem of how to order the training data for a LM. The authors define the notion of a skill, show that skills can be (noisily) identified in training data, show that skills can be ordered, and show that sampling from the data defining a skill in an intelligent manner can improve pretraining and finetuning. They provide both batch and online training algorithms. They present moderately comprehensive experimental results demonstrating the effectiveness of the approach.

Strengths

I really liked this paper. I think it's a great piece of work, and deserves to be published. + The paper addresses an important and timely problem of great interest to a large community + The paper is well-written and polished + The paper strikes an excellent balance of notation, high-level ideas, and solid empirical work + The resulting algorithm seems easy to implement and empirically effective + I enjoyed the derivation of the Skill-IT algorithm - a tidy piece of work. Well done.

Weaknesses

- It's not clear exactly how skills can be effectively identified. Right now, the paper uses various forms of meta data, but this doesn't seem like quite the best (or most general) approach. Maybe something based on the entropy of the text?

Questions

None

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

There is no dedicated section to limitations. I do not foresee any negative societal impacts.

Reviewer paFF7/10 · confidence 4/52023-07-06

Summary

The authors introduce a new way of curriculum learning that is thematically similar to active learning: subsets of the data are broken up into "skills" and in cases where there is a dependency structure of skills, learning precursor skills can help improve performance. When the dependency structure is either complete or empty there is not huge motivation for such approaches (as they layout), but cases where there is a clear dependency structure, they find gains in both final model performance and learning speed. They test GPT-NEO on synthetic tasks that require variable tracking and addition, as well as Natural Instructions. They demonstrates these gains in both pre-training and fine-tuning settings.

Strengths

0. This paper brings together some ideas on how to better train and fine-tune models on related tasks. They introduce a simple but clean (and flexible-enough) formalism of a skill that can be applied in an NLP setting. 1. They demonstrate their method over 3 datasets. 2. If one knows the graph of skills then it seems like this method will tend to be quite helpful in practice. (If not: see my concerns below!). 3. I like Figure 1, making it a tiny bit bigger with bigger fonts would be awesome! I like this paper and I think it could be of practical use to people.

Weaknesses

0. I'd prefer if more space were dedicated to talking about the adjacent matrix. It seems learning it is extremely expensive. How effective is the approximate approach? If we include the learning costs of the adjacency matrix (i.e., give all that extra training to uniform sampling), do the other approaches equal out? I think there are scenarios where we have a clear idea regarding the dependency structure and order. 1. If the adjacency matrix is slightly wrong or noised, how does that impact the results of skill it? How consistent is the matrix over different random seeds / approximations? 2. Continuing on, if there is a large quadratic operation to find the adjacency structure, where each cell basically asks, does learning task i help task j, is there any sense of which this paper is begging the question? I think not, because you're finding a practical and specific way of when and which tasks ia, ib, ic to sample to improve j. Nit: 1. I think the presentation of the LEGO example could possibly be made better, but I don't have great suggestions. 2. The charts are not consistent in sizing. It is tough but doing so would improve presentation.

Questions

0. I think this paper could be greatly improved with some deeper analysis in how the algorithm ends up actually working. Where and how does it differ from a straightforward curriculum? Even for a simple case like addition, are we mostly learning add1 and then add2 and then add3? Is it mixed? Do the proportions shift? Showing diagrams with the proportions + in comparison to random sampling and anti-curriculum could be illuminating. (I now see you have a motivation version of this in Figure 1. I'd really like to see the actual versions. 1. Do you have a sense why Skill-stratified did worse across the board in Figure 6? 2. How can Skill-it be better for Add-1? How is it different than random or curriculum? (To confirm: Those other cases are sampling across all the various Add-#s). 3. Perhaps I missed this, but how does active-learning compare to anti-curriculum?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Seems good.

Reviewer j55A4/10 · confidence 4/52023-07-24

Summary

Review Summary: The paper explores the concept of clustering data into skills to establish a curriculum, aiming to increase the efficiency of language model (LM) training. The concept of a 'skill' is introduced as a unit of behavior with a corresponding data distribution. A skill, in this context, signifies that a model's performance improves on other data that assess a similar skill when it is trained on a dataset associated with that skill. Attempts to use embedding-based clustering for the unsupervised recovery of the curriculum did not yield desired results. However, loss-based clustering proved to be more successful, establishing that validation loss points from the same skill tend to follow similar trajectories. These trajectories were used to cluster by loss per timesteps per run. Equipped with this understanding, the authors suggest it is possible to generate an ordered skill set and extract a curriculum from it. The existence of such an ordered skill set in real data is demonstrated in the paper. The paper also presents 'Skill-It', an approach to dynamically create training datasets, adjusting mixture weights based on expected improvements in generalization on the target task. However, the efficacy of this method is highly contingent on the accuracy of the transfer matrix $A$ approximation. The method employed in the paper consists of two steps: 1. Learning $A$ via brute force or linear approximation, and by approximating $A$ with a smaller model to reduce computational cost. 2. Implementing 'Skill-It' to dynamically sample training data based on the evolution of losses. Using synthetic and real data, the authors show that such ordered skill sets exist and can enable more data-efficient training. However, the paper also acknowledges that these ordered skill sets do not align with intuitive data groupings based on metadata and embedding clustering. Lastly, a novel online data sampling algorithm, SKILL-IT, is introduced for more efficient learning of skills. The authors demonstrate the effectiveness of their approach with improved accuracy in various training scenarios.

Strengths

1. Theoretical Soundness: The paper presents a robust framework based on the innovative concept of 'skills'. The methodology, which encompasses learning an approximation of the transfer matrix $A$, skill-stratified sampling, and the implementation of 'Skill-It', appears to be well-constructed and theoretically sound. 2. Clarity of Presentation: The writing in the paper is clear, making the complex concepts and methodology accessible and comprehensible. The authors effectively communicate their ideas. 3. Potential Impact: If the proposed method is effectively evaluated and shown to significantly surpass existing baselines, this work could potentially make a substantial contribution to the field of language model training. The concepts of ordered skill sets and their application for more efficient, skill-based training could lead to improved performance of language models and open up new avenues for research in this area.

Weaknesses

Methodological Issues: 1. Overlooking Computational Requirements: As the primary goal of this paper revolves around reducing computational expense, it's essential to consider the additional compute requirements of the proposed methods (e.g., graph learning part) in the evaluation of the results. 2. Emphasis on a Singular Experiment: The additional computational requirements could be offset by significant improvements in larger scale models, making the 125M to 1B experiment particularly important. However, this leaves only the results from Figure 7 / Table 8 and Table 9 as practically relevant for real-data experiments. 3. Computational Cost and Hyperparameters: The methodology is not without its complexities, including a fixed computational cost to learn $A$ and the introduction of a new hyperparameter, $\nabla$. These increase the complexity of the method, making it necessary to demonstrate a strong outperformance over the random baseline to justify their inclusion. Experimental Shortcomings: 1. Unspecified Deviations: The authors should clarify whether the deviations reported in the tables are standard errors or standard deviations. Without this clarification, the statistical significance of the results remains uncertain. 2. Missing Data in Section 5.1: The lack of standard deviations or standard errors in the main text necessitates referring to the appendix for assessing the significance of results. For the Addition task, the paper does not sufficiently address the observed equivalence between Skill-It, Curriculum, and Anti-Curriculum. In the case of Natural Instructions (real-data), it's uncertain whether Skill-It significantly outperforms random. 3. Incomplete Reporting in Sections 5.2 and 5.3: The omission of standard errors or deviations in Section 5.2, along with the absence of a random baseline, makes it challenging to evaluate the results. Similarly, for Section 5.3 (Figure 6 / Table 7 and Figure 7 / Table 8), it's unclear whether the results are statistically significant due to the absence of reported standard errors or deviations. 4. Need for Statistical Tests in E.4: As the E.4 experiment is likely the one where the initial cost of computing $A$ can be minimized, it's critical to determine if Skill-It is significantly better than skill-stratified sampling. This requires performing appropriate statistical tests with a sufficient number of trials. 5. Absence of Hyperparameter Ablation for Skill-It: While there's an ablation study for $\nabla$ in E.5, one for Skill-It is notably absent. The authors need to demonstrate that Skill-It isn't hypersensitive to its hyperparameters, or otherwise recognize the associated costs of hyperparameter tuning. 6. Limited Comparison to Stronger Baselines: Even without being experts in curriculum learning, one may wonder why the proposed method isn't benchmarked against stronger baselines from this field. This comparison would be valuable in further establishing the strengths and weaknesses of the proposed method.

Questions

See Weaknesses

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The limitations aren't discussed. I suggest some in the Weaknesses section.

Area Chair 9Cag2023-08-14

Thanks everyone! Reviewer j55A, the author response includes new information about computational costs, hyperparameters, and statistical significance in response to concerns raised by your review. Can you please look over both the global and individual response and let us know whether these concerns have been addressed?

Area Chair 9Cag2023-08-21

Hi reviewer j55A, As we enter the final stage of the review process, it would be very helpful to under whether the rebuttal above addresses your concerns---please let us know!

Reviewer paFF2023-08-18

Thanks for your thorough global (and local!) responses. I am adjusting my review from Weak Accept to Accept. (I had a hard time finding the proportions listed in the appendix... I look forward to seeing the finalized draft!)

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC