The Flan Collection: Designing Data and Methods for Effective Instruction Tuning

Introduction

Large language models such as PaLM (Chowdhery et al., 2022), Chinchilla (Hoffmann et al., 2022), and ChatGPT among others (Brown et al., 2020; Ouyang et al., 2022) have unlocked new capabilities in performing natural language processing (NLP) tasks from reading instructive prompts. Prior art has shown that instruction tuning—finetuning language models on a collection of NLP tasks formatted with instructions—further enhances the ability of language models to perform an unseen task from an instruction (Wei et al., 2021; Sanh et al., 2021; Min et al., 2022).

In this work, we evaluate the methods and results of open sourced instruction generalization efforts, comparing their finetuning techniques and methods. And in particular, we identify and evaluate the critical methodological improvements in the “Flan 2022 Collection”, which is the term we use for the collection of data and methods for data augmentation and instruction tuning, first implemented and used in Chung et al. (2022). Where Chung et al. (2022) focuses on the emergent and state-of-the-art results of combining Flan 2022 with PaLM 540B, this work focuses in on the details of the instruction tuning methods themselves, ablating individual factors, and comparing them directly to prior work by keeping the pretrained model size and checkpoint consistent.

The Flan 2022 Collection offers the most extensive publicly available set of tasks and methods for instruction tuning, which we have compiled in one place. We have also supplemented this with hundreds more of our own high-quality templates, richer formatting patterns, and data augmentations. We show that a model trained on this collection outperforms other public collections on all tested evaluation benchmarks, including the original Flan 2021 (Wei et al., 2021), T0++ (Sanh et al., 2021), Super-Natural Instructions (Wang et al., 2022c), and the concurrent work on OPT-IML (Iyer et al., 2022). As shown in Figure 1, this includes 4.2%+ and 8.5% improvements on the MMLU (Hendrycks et al., 2020) and BIG-Bench Hard (Suzgun et al., 2022) evaluation benchmarks respectively, for equally sized models.

Analysis of the Flan 2022 method suggests the strong results stem both from the larger and more diverse set of tasks, but also from a set of simple finetuning and data augmentation techniques. In particular, training on a mix of examples templatized with zero-shot, few-shot, and chain-of-thought prompts improves performance in every one of these settings, together. For instance, adding just 10% few-shot prompts improves zero-shot prompting results by 2%+. Additionally, enriching task diversity by inverting input-output pairs, as used in (Sanh et al., 2021; Min et al., 2022), along with balancing task sources, are both shown to be critical to performance. The resulting Flan-T5 model converges faster and at a higher performance than T5 models in single-task finetuning—suggesting instruction-tuned models offer a more computationally-efficient starting checkpoint for downstream applications, corroborating Aribandi et al. (2021) and Liu et al. (2022b).

We hope making these findings and resources publicly available will unify resources around instruction tuning and accelerate research into more general-purpose language models. We summarize this work’s core contributions as follows:

Methodological: Show that training with mixed zero- and few-shot prompts yields much better performance in both settings (Section 3.2).

Methodological: Measure and demonstrate the critical techniques to effective instruction tuning: scaling Section 3.3, enriching task variety with input inversion (Section 3.4), adding chain-of-thought training data, and balancing different data sources (Section 3.5).

Results: Demonstrate these technical choices yield 3-17% Held-Out task improvements over existing open source instruction tuning collections (Figure 1).

Results: Demonstrate Flan-T5 serves as a stronger and more computationally-efficient starting checkpoint for single-task finetuning (Section 4).

Open source the new Flan 2022 task collection, templates, and methods for public research.

Public Instruction Tuning Collections

Instruction tuning has emerged as a tool to make large language models (LLMs) and their abilities more useful for interactive dialog and functional tasks. Previous work (Raffel et al., 2020; Liu et al., 2019; Aghajanyan et al., 2021; Aribandi et al., 2021) experimented with large scale multi-task finetuning, to improve downstream single target finetuning, but without instruction prompts. UnifiedQA and others (Khashabi et al., 2020; McCann et al., 2018; Keskar et al., 2019) unified a wide range of NLP tasks into a single generative question answering format, using prompt instructions for multi-task finetuning and evaluation.

Since 2020, several instruction tuning task collections have been released in rapid succession, outlined in Figure 2. Natural Instructions (Mishra et al., 2021), Flan 2021 (Wei et al., 2021), P3 (the Public Pool of Prompts, Bach et al., 2022) aggregated large NLP task collections and templatized them with instructions (zero-shot prompting), specifically for finetuning models to generalize to unseen instructions. MetaICL (Min et al., 2022) also consolidated other task collections (Ye et al., 2021; Khashabi et al., 2020) to train models to learn tasks “in-context” – from several input-output examples, known as few-shot prompting, but in this case without instructions. Each of these works affirmed the scaling benefits of task and template diversity, and some reported strong benefits from inverting the inputs and outputs in templates to produce new tasks (“noisy channel” in Min et al., 2022).

A second wave of instruction tuning collections expanded prior resources: combining more datasets and tasks into one resource, like Super-Natural Instructions (Wang et al., 2022c) or OPT-IML (Iyer et al., 2022), adding multilingual instruction tuning in xP3 (Muennighoff et al., 2022), and Chain-of-Thought training prompts in Flan 2022 (Chung et al., 2022). Both the Flan Collection and OPT-IML contain most tasks represented in prior collections.Note that each work defines datasets, tasks, and task categories differently. For simplicity, we use their own definitions in Section 2. Our work is positioned here, coalescing most of these collections (of collections) and their methods, as the strongest starting point for future open source work.

Concurrent and future work is beginning to explore two new directions: (a) expanding task diversity even more aggressively with synthetic data generation, particularly in creative, and open-ended dialogue (Wang et al., 2022b; Honovich et al., 2022; Ye et al., 2022; Gupta et al., 2022), and (b) offering human feedback signals on model responses (Ouyang et al., 2022; Glaese et al., 2022; Bai et al., 2022a; Nakano et al., 2021; Bai et al., 2022b). We view most of these new directions as likely additive to a foundation of instruction tuning methods.

Instruction tuning on human feedback has demonstrated strong results on open-ended tasks, but at the expense of performance on a wide array of more traditional NLP tasks (Ouyang et al., 2022; Glaese et al., 2022; Bai et al., 2022a; Nakano et al., 2021). (See Ouyang et al. (2022)’s discussion of the “alignment tax”.) Our work focuses specifically on instruction generalization, without human feedback, for two reasons. First, human feedback datasets are far less publicly available than instruction tuning datasets (and may be model-specific). Second, by itself, instruction generalization shows great promise in enhancing human preferred responses on open-ended tasks, as well as improving traditional NLP metrics (Chung et al., 2022). The extent of obtainable progress without expensive human response demonstrations or ratings remains an open question, and an important pursuit to narrow the gap between public and non-public research.

High profile research is increasingly driven by non-public data, as in the case of GPT-3 and others (Ouyang et al., 2022; Glaese et al., 2022). The inaccessibility of these resources inhibits the research community’s ability to analyze and improve these methods in the public domain. We narrow our purview to open source and accessible data collections, motivated by the goal of democratizing accessibility to research.

Flan 2022 Instruction Tuning Experiments

Recent research has yet to coalesce around a unified set of techniques, with different tasks, model sizes, and target input formats all represented. We open source a new collection, first introduced in Chung et al. (2022), denoted “Flan 2022”, which combines Flan 2021, P3++“P3++” is our notation for all datasets in the Public Pool of Prompts (P3): https://huggingface.co/datasets/bigscience/P3, Super-Natural Instructions, with some additional reasoning, dialog, and program synthesis datasets. We defer to Chung et al. (2022) for details of templatization and collection; and in this work we take a deeper look at key methodological improvements and compare the collection on equivalent model sizes to existing collections.

In this section, we evaluate the design decisions in Flan and discuss four in particular that yield strong improvements to the instruction tuning recipe. These design components, outlined in Section 2, are: (I) using mixed zero-shot, few-shot, and Chain-of-Thought templates at training (Section 3.2), (II) scaling T5-sized models to 1800+ tasks (Section 3.3), (III) enriching tasks with input inversion (Section 3.4), and (IV) balancing these task mixtures (Section 3.5). In Section 3.1, we begin by measuring the value of each component and compare the final model against alternative instruction tuning collections (and their methods).

We finetune on the prefix language model adapted T5-LM (Lester et al., 2021), using the XL (3B) size for all models for consistency, unless otherwise stated. While other sizes of Flan-T5 are available, we felt XL was appropriately sized to run large-scale systematic ablations, while being sufficiently large to draw general conclusions. We evaluate on (a) a suite of 8 “Held-In” tasks represented within the 1800+ training task collection (4 question answering and 4 natural language inference validation sets), (b) Chain-of-Thought (CoT) tasks (5 validation sets), and (c) the MMLU (Hendrycks et al., 2020) and BBH (Suzgun et al., 2022) benchmarks as our set of “Held-Out” tasks, as they are not included as part of Flan 2022 finetuning. The Massivley Multitask Language Understanding benchmark (MMLU) broadly tests reasoning and knowledge capacity across 57 tasks in the sciences, social sciences, humanities, business, health, among other subjects. BIG-Bench Hard (BBH) includes 23 challenging tasks from BIG-Bench (Srivastava et al., 2022) where PaLM under-performs human raters. In our ablations, we also evaluate BBH with Chain-of-Thought inputs, following Chung et al. (2022). Additional finetuning and evaluation details are provided in Appendix A.

1 Ablation Studies

Table 1 summarizes the mean contribution to Held-in, Held-out, and Chain-of-thought tasks, by individually deducting methods: mixture weight balancing (“- Mixture Balancing"), Chain-of-thought tasks (“- CoT"), mixed prompt settings (“- Few Shot Templates"), and Input Inversion (“- Input Inversion"). Flan-T5 XL leverages all four of these methods together. We also finetune T5-XL-LM on other collections, including Flan 2021, P3++, Super-Natural Instructions for comparison.

Each of the ablated components of Flan contributes improvements to different metrics: Chain-of-Thought training to Chain-of-Thought evaluation, input inversion to Held-Out evaluations (MMLU and BBH), few-shot prompt training to few-shot evaluations, and mixture balancing to all metrics.

As compared to T5-XL models trained on alternative instruction tuning collections (and their methods), Flan outperforms in almost every setting. While previous collections are tuned specifically to zero-shot prompts, Flan-T5 XL is tuned for either zero- or few-shot prompts. This yields performance margins of +3-10% for most of the zero-shot settings, and margins of 8-17% for the few-shot settings. Most impressively, Flan 2022 outperforms OPT-IML-Max’s much larger (10x) 30B and (58x) 175B models. Next, we isolate some of Flan 2022’s ablated methods individually, to examine the benefits of each.

2 Training with Mixed Prompt Settings

Prior work has shown a wide variety of input templates per task can improve performance. However, separate from the wording of the instruction template, these prior LLMs mostly tune with template sets targeted to a single prompt setting: for zero-shot prompting (Wei et al., 2021; Sanh et al., 2021; Aghajanyan et al., 2021; Aribandi et al., 2021) or for few-shot prompting (Min et al., 2022; Wang et al., 2022c).

An underappreciated design decision in InstructGPT (Ouyang et al., 2022) was to mix training templates for each of these prompt settings, rather than target a single setting. However, since Ouyang et al. (2022) do not examine this choice, we expected a performance trade-off in finetuning for zero-shot or few-shot prompting performance – particularly for smaller models. Instead, we find training with mixed zero- and few-shot prompts significantly improves performance in both settings – most surprisingly, even for models with only 3B parameters.

Figure 3 shows (1) adding as little as 5% few-shot training templates can dramatically improve zero-shot performance, and (2) adding 10%+ of zero-shot data improves few-shot performance too. Both Held-In and Held-Out tasks peak anywhere between 10-90% of few-shot data, but this range is consistently higher than training with only one prompt setting.

3 Scaling Small Models to 1.8k+ Tasks

The most recent and concurrent publicly available instruction tuning efforts, like Flan 2022, train on thousands of tasks (Wang et al., 2022c; Iyer et al., 2022), but operate on different task compositions and underlying training methods. To measure the impact of scaling model sizes and tasks for the Flan 2022 collection, we finetune T5-LM adapted models (Small, Base, Large, XL, XXL) on randomly selected task subsets (8, 25, 50, 100, 200, 400, 800, all 1873). Every finetuning run is guaranteed to include the Held-In tasks, so we can estimate how task scaling impacts the model capacity to maintain performance on a given task its already seen.

Figure 4 demonstrates that both Held-In and Held-Out tasks appear to benefit from adding hundreds of finetuning tasks. Held-in task evaluations peak around 200 total tasks, and diminish in performance as more tasks are added, though larger models peak later and diminish less. Held-out task performance increases log-linearly with the number of tasks, achieving the highest performances with all 1836 tasks. Surprisingly, only T5-Small appears to exceed its Held-Out task performance before 1836 tasks, while larger model sizes continue to improve. These results suggest (a) even T5-Base may not have exhausted its capacity with thousands of tasks, and (b) the largest LMs could benefit from thousands more tasks for Held-In and Held-Out task performance.

One necessary assumption of this analysis is that all tasks are defined and counted equally. Section 3.5 demonstrates how not all task sources are equally beneficial to training, and the model performance may saturate from too many tasks from one source (e.g. Super-Natural Instructions). We would caution conclusions that task scaling beyond 1800 would translate to increased returns without also paying attention to task diversity and quality.

4 Task Enrichment with Input Inversion

Prior instruction tuning work has enriched their diversity of tasks by inverting the ($x$, $y$) input-output pairs in supervised tasks—referred to as “prompts not intended for the original task” in P3 (Bach et al., 2022) or the “noisy channel” in MetaICL (Min et al., 2022). For example, a dataset may be originally designed for, given a question $x$, evaluate if a model can answer $y$. Input inversion instead gives a model the answer $y$ and trains it to generate the question $x$. This is an easy method to enrich the task variety given a limited set of data sources. However, it isn’t clear that this method remains helpful when 100s of unique data sources and 1000s of tasks are already available.

To assess this, we enrich our mixtures with input inverted tasks (details and examples in Appendix B) and measure the effect. In Table 1 we find this is not beneficial for Held-In performance, but strongly beneficial for Held-Out performance. These benefits invigorate the prospect of data augmentation techniques for LLM finetuning, which had previously been shown to have diminishing returns the longer models are pretrained (Longpre et al., 2020).

5 Balancing Data Sources

Scaling architecture size and the number of tasks are effective, but our results suggest the mixture weighting deserves as much attention to optimize results. To converge on a balanced weighting, we omit different sets of task sources, one at a time (Flan 2021, T0-SF, Super-Natural Instructions, Chain-of-Thought, Dialog, and Program Synthesis), and rank their contributions on the MMLU benchmark.Following Chung et al. (2022) we refer to the subset of P3++ that is not in Flan 2021 as T0-SF (SF stands for “sans Flan”)..

As shown in Table 2, Flan 2021 and T0-SF are among the most beneficial mixtures, followed by Super-Natural Instructions and Chain-of-Thought, with Dialog and Program Synthesis last. These findings are corroborated by Iyer et al. (2022) who extensively test data mixing proportions, and also determine their Flan 2021, T0-SF, and T5 mixtures are the most broadly beneficial. Additionally, they find Super-Natural Instructions has limited scaling benefits on Held-Out task performance, which they relate to its unique input format and instruction design. Notably, Chain-of-thought finetuning appears beneficial across all our evaluation settings, especially considering they contain far fewer tasks than Flan 2021, T0-SF or Natural Instructions.

We used these findings to significantly narrow the mixture weights search space, and used our practitioner’s intuition from there. This strategy is simple but effective, as shown in Table 1, but leaves ample room for more sophisticated future work.

6 Discussion

OPT-IML (Iyer et al., 2022) presents the closest comparison to this work, including a similar collection of tasks, examples and techniques. However, while their used tasks are all publicly sourced, their collection, with templates, processing, and example mixing, is not released, and as a result cannot be easily compared. Iyer et al. (2022) report that Flan-T5-XL (3B) and XXL (11B) outperforms OPT-IML-Max 175B on both MMLU and BBH. As they discuss, these differences may arise from any combination of pre-training, model architecture, and instruction tuning. Model architecture and pretraining before instruction tuning can play a significant role (Wang et al., 2022a). But there are many other details in instruction tuning that may vary between Flan 2022 and OPT-IML. Likely candidates are are: example templatization, how the mixed input prompting procedures are used at training, and task composition.

How significant are each of these difference? While OPT-IML contains more tasks than Flan 2022, we estimate approximately $94\%(2067/2207)$ are also used in the Flan 2022 collectionThis is calculated using their definition of “task” (reported in Iyer et al. (2022)’s Table 1), which does not deduplicate across collections., and very few tasks in Flan 2022 are not contained in some format in OPT-IML. This suggests the overall difference in task diversity is not significant when using a shared definition of “task”. Task mixture rates also emphasize similar sources, including Flan 2021 (46% vs 20%), PromptSource/P3 (28% vs 45%), and Super-Natural Instructions (25% vs 25%), for Flan 2022 and OPT-IML respectively.Note that 46% weight for Flan 2022 is actually on Muffin from Chung et al. (2022) which combines Flan 2021 with new dialog and program synthesis tasks. OPT-IML’s other collections (Crossfit, ExMix, T5, U-SKG) are not weighted significantly: 4%, 2%, 2%, 2% respectively.

We believe example templatization and the mixed prompt formats may pose the largest differences with OPT-IMLs instruction tuning. Our template repository was significantly updated from Flan 2021, adding variety not just in instructions, but also along dimensions. For instance, the templatization procedure varies where the instruction is placed (before or after few-shot prompts), the spacing and separators between few-shot and Chain-of-Thought prompts, and the formatting permutations of answer options (and their targets) for multiple-choice examples, which sometimes includes and sometimes excludes answer options in the inputs or exemplars. While we do not have dedicated experiments comparing many iterations of development, we found these procedures dramatically augment input variety and showed repeated performance improvements. Our example templatizing procedure is open sourced for inspection and future work.

Instruction Tuning Enhances Single-Task Finetuning

In applied settings, machine learning practitioners deploy NLP models finetuned (FT) specifically for a single target task, usually where finetuning data is already available. While prior work has shown the benefits of intermediate finetuning (Pruksachatkun et al., 2020; Vu et al., 2020) or multi-task finetuning (Aghajanyan et al., 2021; Aribandi et al., 2021) for downstream tasks, this has not been studied extensively for instruction-tuned models.

We evaluate Flan 2022 instruction tuning as an intermediary step before single target finetuning, to understand if Flan-T5 would serve as a better starting checkpoint for applied practitioners. We evaluate three settings in Figure 5: finetuning T5 directly on the target task as the conventional baseline (blue bars), using Flan-T5 without further finetuning (beige bars), and finetuning Flan-T5 further on the target task (red bars).

For both sets of Held-In and Held-Out tasks examined, finetuning Flan-T5 offers a pareto improvement over finetuning T5 directly. In some instances, usually where finetuning data is limited for a task, Flan-T5 without further finetuning outperforms T5 with task finetuning.

Using Flan-T5 as a starting checkpoint has an added benefit in training efficiency. As demonstrated in Figure 6, Flan-T5 converges much more quickly than T5 during single target finetuning, as well as peaking at higher accuracies. These convergence results also suggest there are strong green-AI incentives for the NLP community to adopt instruction-tuned models, like Flan-T5 for single-task finetuning, rather than conventional non-instruction-tuned models. While instruction tuning is more computationally-expensive than single-task finetuning, it is a one-time cost. On the contrary, pretrained models that require extensive finetuning become more costly when aggregating over many millions of additional training steps (Wu et al., 2022; Bommasani et al., 2021). Instruction-tuned models offer a promising solution to significantly reduce the amount of finetuning steps across a wide swathe of tasks, if they are adopted as a new standard starting point for single-task finetuning.

Related Work

As the foundation of instruction tuning, the practice of pretraining one general-purpose language representation that is useful for multiple downstream tasks has a long tradition that goes back at least Mikolov et al. (2013) and Dai and Le (2015). In 2018, Peters et al. (2018) and Devlin et al. (2019) cemented the paradigm of pretraining a large model on a large unsupervised corpus, and the field of NLP quickly converged to using these models which substantially outperform the prior art of non-pretrained task-specific LSTM models on all tasks. However, the dominate way to access that high-quality syntactic and semantic knowledge encoded in pretrained models was not to prompt them with instructions, but to train an additional task-specific linear layer that maps the model activations into numerical class labels. A short year later, Radford et al. (2019), Raffel et al. (2020), and Lewis et al. (2020) popularized the notion that downstream tasks—and multiple tasks—can be jointly learned by directly using the pretrained LM head to generate the answers in natural language (cf. task-specific numerical class labels), the task-general nature of these generative models became the precursor to many multitask transfer learning studies (McCann et al., 2018; Khashabi et al., 2020; Ye et al., 2021; Vu et al., 2020), which in turn led to the first wave of instruction tuning as described in Section 2.

The continuing advancement in research on the pretraining corpora, architectures and pretraining objectives of LMs also has a large impact on instruction tuning. As of 2022, decoder-only left-to-right causal Transformers dominate the market of models larger than 100B (Brown et al., 2020; Thoppilan et al., 2022; Rae et al., 2021; Chowdhery et al., 2022; Hoffmann et al., 2022), and all models of such size class with fully public model parameters are decoder-only (Wang and Komatsuzaki, 2021; Le Scao et al., 2022; Zhang et al., 2022), the decision of which are often due to better hardware and software framework support. However, Raffel et al. (2020), Lewis et al. (2020), and Tay et al. (2022a) have consistently found that left-to-right causal language modeling is a suboptimal objective, while Tay et al. (2022b) and Wang et al. (2022a) particularly showed that a mixture of non-sequential objectives is much superior for downstream tasks with zero-shot and few-shot prompting. An additional factor which remains under-explored is the relationship between pretraining corpora, instruction tuning, and downstream abilities. Typically, public models are all trained on one of a few public corpora: C4 (Raffel et al., 2020), The Pile (Gao et al., 2020), or ROOTs (Laurençon et al., 2022).

In Section 2 we outline major developments in instruction tuning. Other important developments include the prospect of complimenting or replacing few-shot in-context learning-the currently predominate method of evaluating pretrained and instruction-tuned models—with parameter-efficient tuning. As standard finetuning of models larger than 100B requires a high number of accelerators with the right interconnects often too expensive even for many industry labs, parameter-efficient tuning (a.k.a. continuous or soft “prompt tuning”) shows that only updating a small subset of model parameters can reach comparable performance as fully tuning all model parameters (Lester et al., 2021; Vu et al., 2022; Hu et al., 2021; see He et al., 2022 for a detailed analysis). Notably, Liu et al. (2022b) show that, due to the long sequence length of few-shot ICL and that the few-shot exemplars need to be repeatedly inferenced for evaluating every example, parameter-efficient tuning can be computationally cheaper and higher performing than in-context learning. Further, Liu et al. (2022b), Vu et al. (2022), Wei et al. (2021), and Singhal et al. (2022) collectively show that both single-task and multi-task parameter-efficient tuning can be productively combined with instruction tuning, either before or after regular full-model instruction tuning. This line of work makes it easy for other researchers to build on top of a general-domain instruction-tuned model, and collect a custom instruction-tuning mixture for their use, e.g., with multiple modalities (Ahn et al., 2022; Huang et al., 2022; Xu et al., 2022) or special domains such as science and medicine (Lewkowycz et al., 2022; Singhal et al., 2022).

Instruction tuning is part of a line of work designed to “align” language models with more useful objectives and human preferences. In the absence of such methods, language models are known to demonstrate toxic/harmful behaviour (Sheng et al., 2019; Liang et al., 2021; Wallace et al., 2019), generate non-factual information (Maynez et al., 2020; Longpre et al., 2021; Devaraj et al., 2022), and other challenges in deployment and evaluation (Zellers et al., 2019; McGuffie and Newhouse, 2020; Talat et al., 2022). Analyzing, evaluating and mitigating these problems pose a promising direction for future work (Gao et al., 2022; Ganguli et al., 2022). Instruction tuning warrants greater investigation, as it has already demonstrated itself an encouraging remedy in reducing NLP bias metrics, as shown in Chung et al. (2022).

Conclusions

The new Flan 2022 instruction tuning collection unifies the most popular prior public collections and their methods, while adding new templates and simple improvements like training with mixed prompt settings. The resulting collection outperforms Flan 2021, P3++, Super-Natural Instructions, and OPT-IML-Max 175B on Held-In QA, NLI, and Chain-of-Thought tasks, and Held-Out MMLU and BBH, often by large margins. Results suggest this new collection serves as a more competitive starting point for researchers and practitioners interested in both generalizing to new instructions, or finetuning on a single new task.

Acknowledgements

We would like to thank Ed H Chi, Xinyun Chen, and Colin Raffel for their advice and feedback on the paper.

References

Appendix

The Flan Collection experiments are assembled and run using T5X (Roberts et al., 2022). Our instruction tuning follows the same setup described in Chung et al. (2022). For few-shot and few-shot Chain-of-Thought prompts during finetuning our templatizing procedure generates few-shot examples with 2, 3, or 5 exemplars. The experiments in this work use a slightly earlier version of the Flan 2022 collection the one we are releasing, which had some minor improvements to the templates.

The mixture weights used to balance the various sources of data were informed by experiments in Section 3.5, along with the resulting practitioner intuition.

A.2 Single-Task Finetuning

For single-task finetuning, described in Section 4, our models are finetuned for 100,000 steps for all tasks. We use a constant learning rate of 0.001, a dropout probability of 0.1, and a batch size of 128 length-512 sequences. We save a checkpoint every 20 steps and report test performance on the model checkpoint corresponding to the highest validation performance. For tasks without a validation split, we hold out 1024 training examples for validation. For tasks without a test split, we hold out 1024 training examples for validation and report results on the original validation set. For PubmedQA, we do not use any of the unlabeled and artificially generated QA instances associated with the dataset. For CxC, we only consider the text-text portion of the dataset, following Vu et al. (2022). For tasks with less than 1K training examples, we report average results across 3 random seeds.

We also evaluate on certain metrics to account for label skew in some of the datasets, as shown in Table 3.

A.3 Evaluation

For Held-In evaluations we use the validation sets from 4 question answering (QA) tasks, BoolQ, ARC Easy, ARC Challenge, and AI2’s Middle School Science Exams, and 4 natural language inference (NLI) tasks, including ANLI R1, R2, R3, and RTE. These datasets are contained in the Flan 2022 finetuning collection and represent challenging benchmarks, often used to evaluate LLMs on QA and NLI. The Held-In score is the mean accuracy across these 8 tasks.

For the Chain-of-Thought (CoT) evaluation, we use the mean accuracy across 5 datasets which have been prepared with prompts which request step-by-step explanations in their target answers: GSM8K, StrategyQA, SVAMP, Asdiv, and CommonsenseQA.

For the Held-Out tasks, we use MMLU’s suite of 57 exams, and BBH’s suite of 23 tasks where PaLM performed worse than the average human annotators. MMLU tasks were removed from the Super-Natural Instructions part of the Flan 2022 collection at training, to ensure they were Held-Out.

Appendix B Input Inversion Details

For the input inversion experiments we note that Flan 2021, P3++, and Super-Natural Instructions already implicitly include tasks that have been inverted, e.g. question answering to question or context generation. Consequently, we choose to also create input inversions for the remaining datasets in the Flan 2022 collection, including for the Dialog, Program Synthesis, and Chain-of-Thought tasks.

As examples: for Dialog tasks, we write template instructions asking for the previous conversational history from the current dialog turn; for program synthesis we ask for the coding question which the code solves; and for Chain-of-Thought we include every permutation of the query-answer-explanation triple, where at least one of the three appears as the in output. An illustration of Chain-of-Thought input inversion permutations are shown in Figure 7.

These inversions are mixed in with the existing tasks at a rate of 30%, meaning for a Dialog task, 3 inverted examples will be generated for every 10 regular examples. We choose this rate for simplicity, approximately mirroring prior work, and leave the large space of exploration for future work.