MUFFIN: Curating Multi-Faceted Instructions for Improving Instruction Following

We are grateful for your detailed review and constructive comments! We're pleased you found our work novel, the proposed technique strong, and the experiments solid.

Some of your suggestions are helpful and have been taken by us seriously. In this response, we add more experiments and try to address your concerns one by one:

---

Q1: Are there any specific mechanisms to guarantee the trustworthiness of the facets and output generation?

Thanks for pointing this concern out.

1. First, for the output annotation, we do admit that there is no specific mechanism to guarantee trustworthiness (except human verification), which is a common headache that almost all of the previous works suffered from. However, since the main objective of this paper is to prove the effectiveness of the proposed paradigm instead of simply achieving SOTA with a higher-quality dataset, we try to avoid any sophisticated strategies to cherry-pick training instances (e.g., using a small neural model to help filter hallucinated outputs). But of course, your suggestion is highly appreciated; it shall be our next work to release an even higher-quality dataset with the proposed paradigm (e.g., human-annotated Muffin-v2).

2. As for the facets generation, since it is one of the unique designs and critical steps for creating the dataset with the proposed paradigm (i.e., Scaling Task per Input), we have some simple yet effective operations in our framework to improve its reliability:

   • Explicit Prompting: we added some explicit constraints in the prompt of facets generation, to ensure the quality and diversity of the facets, e.g., “each facet should clearly describe one specific textual attribute…” and “instructions should strictly be based on the facet…” (pls see Table 4 and Table 5 in Appendix C).
   • Joint Filtering: since the generated facets are eventually used for creating instructions, the proposed “Instruction Filtering” can also jointly filter those hallucinated facets that may lead to unanswerable instructions (pls see the second paragraph at page 4).

   We proved the effectiveness of the above methods in Q2.

---

Q2: This work lacks a clear big picture of the generated facets (the reliability and diversity of facets should be quantified); showing the facets with fine-grained categories may help.

We highly agreed with your concern and realized the importance of showing the quantification of generated facets. We carefully took your comments and tried to address your concern by elaborating on the following three dimensions: statistics, quality, and diversity.

---

1. Facets Statistics.

We hope the most straightforward way to show a big picture of the generated facets is to illustrate its statistics. As shown below, similar to Table 9 in our paper, we report additional statistics related to facets (the statistics after conducting “Instruction Filtering”).

---

2. Factes Quality.

To estimate the quality and reliability of the generated facets, and also to demonstrate the effectiveness of our facets-controlling methods (in Q1), we conducted further human verifications on the facets.

Specifically, we randomly collected 100 inputs from Muffin, along with their corresponding facets (889 facets in total) and instructions (2,265 instructions in total). Then, we asked an experienced human annotator to evaluate the following two metrics:

• Input-to-facet correctness: whether each facet correctly described the given input (reflecting the quality and reliability of facts).
• Facet-to-instruction correctness: whether the subsequent instructions are reasonably related to the given facet (reflecting the utility of facets).

The final results are shown below:

These results demonstrate the good quality and utility of the facets generated by our method. It's also worth noting that, after we conducted "Instruction Filtering", the input-to-facets correctness increased from 83.28% to 90.78% (the score above), proving the effectiveness of "Instruction Filtering" that can jointly filter a considerable amount of hallucinated facets.

---

3. Facets Diversity.

As for the diversity, we are sincerely grateful for your suggestions. Inspired by your comments, we asked the human annotator to discover the categories from the facets (i.e., clustering instead of classification), because it's hard to predefine the categories that may introduce personal bias.

We used the same instances as in the previous analysis (100 inputs and 889 facets). The human annotator was asked to summarize a "keyword" (category) for each LLM-generated facet. If the current category has already appeared before, the annotator also has to "cluster" those facets that belong to the same category together. In doing so, more and more novel categories can be discovered.

We then calculated the following two metrics:

• Intra-diversity: how many unique categories there were for each input’s facets, and reported an averaged ratio among all the inputs.
• Inter-diversity: overall categories distribution among all the 889 facets, namely the global frequency of each unique facet.

We found that the generated facets achieve 91.55% intra-diversity, meaning the facets of the same input are pretty diverse.

Meanwhile, there are a total of 276 unique categories out of the 889 facets, which implies a high inter-diversity of the facets. The distribution plot of inter-diversity with detailed categories can be found at this Anonymous Link.

Note that the human-annotated facet categories cover a wide range of topics. Some facet categories may describe the structural attributes of the input (e.g., length, conversation format), while others may focus on the specific textual content (e.g., the input contains a location or time entity).

We will update these analyses in our next version and will try to elaborate on more details.

Q5: Comparing this work with the previous works (e.g., self-instruct), it's hard to identify the differences.

We highlight the differences between this work and the previous works below:

• Motivation.

The main motivation of almost all previous works is the same — how to use LLMs to collect a large-scale instruction-following dataset automatically.

However, the essential motivation of this work is totally different — instead of simply bootstrapping more data, our research question is how to reformulate the current learning paradigm for a better instruction-following capacity of LMs.

We first thought deeply about the scaling-input paradigm; it follows a conventional multi-task learning scheme, where each task instruction has multiple input-output pairs. We assumed that there is a potential harm: models are conditioned on the same instruction but different inputs to generate different outputs, which implicitly teaches models to focus on inputs rather than instruction (because the inputs seem to be the key information to generate correct outputs in this case).

Naturally, we came up with the proposed paradigm, where models are conditioned on different instructions but the same input to generate totally different outputs.

In a word, we respectfully disagree with the opinion that the proposed paradigm is merely a simple incremental production from the previous work. In contrast, we believe it aligns well with human intuition and can inspire future work in this area.

• Method

Following the aforementioned motivation, our data curation method is also quite different from the previous works due to the unique challenge in our paradigm.

Most of the previous works tried to use existing instructions (I) as demonstrations and asked LLMs to brainstorm new instructions (I’), then created the subsequent input (X) and output (Y). The whole procedure can be succinctly described as I ⇒ I’ ⇒ (X,Y).

Contrastingly, we start our data curation procedure from the input text (X) and try to gather diverse instructions (I’) for it, and then annotate the outputs (Y), namely X ⇒ I’ ⇒ Y.

In addition to the different procedures, creating instructions from the input (X ⇒ I’) is much more challenging than generating from existing instructions (I ⇒ I’), because the former requires more restrictions (instructions should “match” the input, there is less freedom) and is also much harder to be controlled (due to the hallucination of the LLMs). That’s why we proposed two different methods for curating enough high-quality instructions, and used input-oriented facets to guide the instruction brainstorming.

To the best of our knowledge, due to the unique nature of the proposed paradigm, collecting the corresponding dataset requires more special treats and designs in the method, which any previous work has never covered.

---

Q6: Share the limitations of this work.

We are willing to shed light on the following two limitations of this work:

1. Since our initial motivation is to drive LMs to learn to follow the instructions instead of mapping inputs to outputs (as mentioned before), the contribution of this work mainly focuses on the changes in the dataset paradigms. However, we still follow a traditional supervised fine-tuning scheme to train the LMs. Therefore, it remains unclear if our paradigm shift can really make a difference during the learning procedure (such as the difference in gradient). But of course, it's sometimes difficult for humans to interpret the "learning of models". Therefore, a special learning objective designed for “Scaling Task per Input” may help to enhance our conclusion.

2. What's more, this work is still following a dataset-scaling scheme in this area; considering the recent findings on "less is more" [3], we think it's also worthwhile to investigate whether we do need to "scaling task per input" or just try to sample "less task per input" and achieves more efficient instruction learning.

Regarding the above limitations, there could be considerable space in this area for the future to dig deep into.

---

References:

[1]. Self-Instruct: Aligning Language Models with Self-Generated Instructions. (ACL 2023)

[2]. How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources. (NeurIPS 2023 Datasets and Benchmarks).

[3]. Lima: Less is more for alignment. (ArXiv 2023)

Dear Reviewer#4 (od6V),

---

We are writing to inquire if you have had the chance to review our detailed response to your comments. We greatly value your feedback and would appreciate any further questions or comments you might have.

---

Sincerely,

Authors of Paper#6820

Your comments are very much appreciated! We are glad you found our idea interesting and the presentation clear. We took your comments carefully and have added some missing experiments you mentioned. We address your concerns one by one as follows:

---

Q1: Muffin highly relies on the human-crafted SuperNI dataset.

Our data curation method does utilize the resources from SuperNI, but it’s hard to say our method “highly” relies on SuperNI, because:

• For “Instruction Brainstorm”, we only used 3 instructions from SuperNI as demonstrations (which is similar to all the previous works); For “Instruction Rematching”, we extracted instructions and inputs from SuperNI separately while NOT keeping the original correspondence between them (the instruction-input correspondence should be the most critical part of human annotation).
• Additionally, our method can be practically applied to any textual resources and still results in comparable instruction-following performances (please refer to the experimental results in Q3).

---

Q2: Muffin is sourced from SuperNI, but introduces more noise.

We agree with your concern and address it with the following points:

• To our knowledge, all the LLM-generated datasets (e.g., alpaca, self-inst) introduced more noise compared with the source datasets. However, one main contribution of this area is to use LLMs to automatically curate instruction-following instances instead of relying on expensive human labor.
• According to our human analysis, Muffin suffers less noise than previous works (pls see Figure 4 in Appendix E).

---

Q3: The authors should create new instructions on the inputs sourced from Dynosaur or Unnatural Inst. to ensure a fair comparison.

This is a very good suggestion!

During this discussion period, we gathered the inputs from other LLM-generated datasets (including Dynosaur and Unnatural-Inst, as suggested by you), and applied our “Instruction Brainstorm” methods to them. Finally, we created two small-scale datasets (about 16k), namely Muffin-Dynosaur and Muffin-Unnatural, and compared their performances with Dynosaur and Unnatural-Inst (in the same size).

The results (with T5-3B models) can be found in the following table:

---

We observed that Muffin consistently demonstrates superior generalization performance, regardless of the text sources employed. It further suggests that Muffin's robust ability to follow instructions is not reliant on the input resources but rather stems more from our crucial contribution—the diversified instructions per input. Therefore, we anticipate that our method can be ideally applied to any resources while still crafting diverse task instructions.

---

Q4: Is it worth trying this paradigm or simply using the queries to chatgpt to collect more data?

We hope we can answer your question by plotting the scaling trends of different datasets. Specifically, we randomly sampled subsets from the original datasets and trained T5-3B on them to show the trends (10%, 30%, 50%, 80%, 100%).

The resulting figure can be found in this Anonymous Link.

In the range of 68K, MUFFIN exceeds the baselines by a noteworthy margin (average scores on 4 evaluation benchmarks). Other baselines may only be comparable to our data results when they continue to be scaled to several times the size of our data.

More importantly, the performances of some datasets even decrease after scaling to a larger size (perhaps due to the noise in these LLM-generated datasets), such as Self-Inst. and Dynosaur. Therefore, we conjecture that our paradigm is more efficient than simply collecting more data from the other paradigms.

---

Dear Reviewer#2 (cEte),

---

We are writing to inquire if you have had the chance to review our detailed response to your comments. We greatly value your feedback and would appreciate any further questions or comments you might have.

---

Sincerely,

Authors of Paper#6820

Thanks for your efforts in reviewing! We are happy you found our paper well-written, the proposed method reasonable, and the empirical results solid.

Some of your suggestions are very important to improve our work. We adopted your comments and tried adding more analyses during this discussion period.

As shown below, we'd like to address your concerns in detail.

---

Q1: A better ablation on the trade-off between diversity and quality.

This is a great point for better evaluating the proposed dataset. Your suggestion on the ablation between “facets-based instruction brainstorm” and “GPT-4 based rematching” is a good perspective to investigate this question.

We put the ablation results in Table 13 in Appendix M, where we tested the ablation performances on the development set of SuperNI to avoid cherry-picking (same as the previous work [4]). Here, for your convenience, we also put that table below.

We found that both these methods contribute to the final performance, and seem to show complementary effects on each other (combining the two methods resulting in better performance).

This conclusion aligns with previous works [1][2], that both factors matter greatly in the final generalization of LMs. However, to our knowledge, it’s still hard to quantify their exact weights because various variables can affect the conclusion (e.g., the base models and the downstream generalization tasks).

We will shed more light on this question in our next version.

---

Q2: The paper lacks scaling analysis of the datasets.

Thanks for your reasonable suggestions. It is critical to investigate the scaling trends to prove the robustness of our dataset.

We plotted the performance trends of different datasets (including Muffin and those competitive baselines) with various scales (10%, 30%, 50%, 80% 100%).

For your convenience, you can find the resulting figure at this Anonymous Link.

The results demonstrate that Muffin consistently achieves superior performances among various scales, and exhibits a stable performance boosting w.r.t. the scaling. However, for some other baseline datasets, the performance even decreased after a certain scale threshold (perhaps due to the effect of the noise).

---

Q3: Insufficient discussion on the task leakage.

We do agree with your concern about the task leakage analysis.

To our knowledge, since there is no SOTA leakage detection method specifically designed for instruction tuning, we followed the previous work [3] using ChatGPT to judge the task leakage of different training datasets.

Given an instance from a training dataset, we asked ChatGPT whether it belongs to any task categories from the evaluation benchmark. For example, the SuperNI-Test has 12 task categories, and we requested ChatGPT to perform a 13-way classification (one “None of above” label) on the training instances. Finally, we report how many instances from this training dataset contain task leakage.

The following table illustrates the results, comparing Muffin with the most competitive baseline datasets. Though there are some shifts in the leakage ranking compared with our original analysis (Figure 6 in Appendix N), the overall conclusion is still the same — Muffin demonstrates relatively low evaluation task leakage across all four benchmarks.

The main objective of our analyses is to address the impact of other factors that may affect the contribution of our proposed paradigm. Though these leakage analyses are not that perfect, we believe the task leakage is not a serious issue affecting our conclusion, according to the above consistent observations.

---

Q4: Particular reasons that different GPT models are selected for different purposes.

During our preliminary trials, we observed GPT -4's superiority over ChatGPT in "Instruction Rematching", which can produce much more precise pair matching than ChatGPT. While for "Instruction Brainstorm", we found that ChatGPT was good enough and much cheaper. That's why we use ChatGPT for "Instruction Brainstorm" and GPT-4 for "Instruction Rematching". We will highlight this point in our next version.

---

Dear Reviewer#3 (j2qL),

---

We are writing to inquire if you have had the chance to review our detailed response to your comments. We greatly value your feedback and would appreciate any further questions or comments you might have.

---

Sincerely,

Authors of Paper#6820

Many thanks for your detailed review! We are glad you found this work interesting, the experiments solid, and the presentation clear. During this discussion period, we conducted more experiments per your suggestions to improve this work. As shown below, we summarize your main concerns and address them one by one:

---

Q1: Whether a hybrid paradigm that scales both (Scaling Inputs & Scaling Tasks per Input) according to some mixture proportion would be most effective.

This is a great point to investigate!

To figure out the answer to this question, we mixed Muffin (Scaling Tasks per Input) with SuperNI-Train (Scaling Inputs) with various proportions (0.0, 0.1, 0.2, …, 0.9, 1.0), where the proportion means how many instances are from Muffin. Then, we trained T5-3B on these mixtures and reported the average performances on the four benchmarks.

For your convenience, we plotted the figure at this Anonymous Link.

Note that, SuperNI-Train is created by humans and has much higher quality. That’s why using only SuperNI-Train resulted in the best performance in this figure (i.e., proportion == 0).

However, the most interesting finding is that, after mixing Muffin with more high-quality data from SuperNI-Train, the model’s performance first dropped and then increased.

Thus, we draw mainly two conclusions here:

1. The dataset paradigm does affect the learning efficiency. Different dataset paradigms do have obviously various impacts on the model’s performance. Therefore, simply combining two different paradigms can even hurt the model’s instruction-tuning efficiency, meaning the paradigm is a critical factor.
2. The superiority of the proposed paradigm. Meanwhile, the effect of dataset paradigms is even greater than that of data quality (adding a small proportion of high-quality data from other paradigms even harms Muffin’s performance), further implying the proposed paradigm's effectiveness.

This observation also inspires us to release an even higher-quality dataset in the future, that has both quality and paradigm superiority (as also suggested by reviewer#4).

---

Q2: The performances with the most recent LLMs (e.g., LLaMA).

We agree that we have to adopt some other models besides T5.

During this discussion period, we conducted experiments with the most advanced decoder-only LMs, namely LLaMA-2 (13B), to further enhance the reliability of our conclusion. Since the training and inference of LLaMA-2-13B were time-consuming, we reported only the performances of Muffin and those most competitive baselines; meanwhile, we randomly selected 2k testing instances from each evaluation benchmark (every evaluation task was covered) to show the comparison, instead of the whole testing benchmarks (which will take more than two weeks).

The results are shown in the following table. The main conclusion is similar to T5’s results in our paper, where Muffin consistently gains better overall performances compared with the baselines.

We will add the completed inference results of LLaMA-2-13B in our next version, and fill in the performances of the "original models" trained on those baseline datasets to the "Existing Systems" in Table 1.

---

Q3: The presentation of human evaluation omits key details.

We are sorry for letting you get confused.

• For the models used in human evaluation: We use T5-3B through all the analysis sections. We will highlight it by mentioning it in the table captions.
• What “input” was provided to the human when facing "input-free" tasks: For those "input-free" tasks, we just left the "input" field empty (pls refer to Table 12 in Appendix L). For a better illustration, we provide the figure of the human evaluation interface at this Anonymous Link.

---

Dear Reviewer#1 (BsoU),

---

We are writing to inquire if you have had the chance to review our detailed response to your comments. We greatly value your feedback and would appreciate any further questions or comments you might have.

---

Sincerely,

Authors of Paper#6820

Dear Reviewer#1 (BsoU),

---

Thanks so much for reading our response and raising the score. It's our pleasure that our additional experiments and efforts can earn your recognition.

Regarding the interesting results of mixing SuperNI, we will definitely attach it with more details and try to figure out more reasonable explanations in our next version.

Thanks again for your recognition of our paper's contributions.

---

Best regards,

Authors of Paper#6820