Inverse Scaling Prize

TL;DR: Win up to $100,000 for finding an important task where larger language models do worse.

Submissions due August 27, 2022 (Round 1) and October 27, 2022 (Round 2).

The contest has ended! Results: Round 1, Round 2.

Recent changes

11 October, 2023

• Added Modus Tollens caveat to data-release README

21 March, 2023

• Updated prize pool info

1 March, 2023

• Released data for all winning tasks

17 December, 2022

• Updated prize eligibility for FAR employees

12 December, 2022

• Added prize terms update to the ‘Prize information’ section
• Updated ‘About us’

9 October, 2022

• Added Huggingface Hub evaluation setup description to the tips section

4 October, 2022

• BUG FIX: Reported total probabilities should now be more accurate for all classification tasks

26 September, 2022

• Demonstrating positive scaling on the ‘incorrect’ answer is now allowed
• Added stronger recommendation to aim for roughly 1000 examples
• Added requirement to name the task
• Added request to include data for control experiments
• Added field to specify how the data was generated
• Added field for links to dataset sources
• Added field for code that generated dataset
• Added requirement that submissions in multiple parts should upload all .csv files together in one .zip
• Added request to make file names anonymous
• Added option to use a variable number of classes in classification datasets
• Added print out to colabs of the total probability given to class labels
• Added reminder that the submitted plot should be from our official colab
• Added request that people edit their form submission rather than resubmit to update
• Added reminder to specify correct behavior on the task
• Added field to specify whether the task is zero-shot or few-shot
• Updated terms and conditions

Motivation

As language models get larger, they seem to only get better. Larger language models score better on benchmarks and unlock new capabilities like arithmetic [1], few-shot learning [1], and multi-step reasoning [2]. However, language models are not without flaws, exhibiting many biases [3] and producing plausible misinformation [4]. The purpose of this contest is to find evidence for a stronger failure mode: tasks where language models get worse as they become better at language modeling (next word prediction).

The standard paradigm in natural language processing today is to pretrain large language models to autocomplete text corpora. The resulting models are then either frozen and used directly for other tasks (zero-shot or using few-shot learning), or additionally trained on other tasks (fine-tuning). We focus on the case of zero-shot/few-shot evaluation on downstream tasks without task-specific gradient optimization: it's typically easier to use in practice and to study.

Scaling laws [5][6] show that language models get predictably better (in terms of test loss and downstream performance [7]) as the number of parameters, amount of compute used, and dataset size increase. The improvement follows a power law in each of parameters, compute, and dataset size. We hypothesize that there are tasks with trends in the opposite direction: task performance gets monotonically, predictably worse as the overall test loss of the language model improves. We call this phenomenon inverse scaling, in contrast with the standard scaling laws. There are some tasks that appear to show inverse scaling under some conditions [4][8][10], but such tasks appear to be rare.

This contest aims to find inverse scaling tasks, especially those of importance to the safe and responsible use of language models. We hope that task submissions will teach us more about what types of tasks exhibit inverse scaling; inverse scaling tasks will also highlight potential issues with the current paradigm of language model pretraining and scaling. Inverse scaling tasks are important because they represent a mismatch between the behavior we want language models to exhibit and the behavior we get in practice from the training objectives and data we use. As language models continue to get bigger and used in more real-world applications, it is important that they are not increasingly getting worse or harming users in yet-undetected ways.

After two rounds of the contest, we will write a survey of the submitted tasks and other examples found in the literature. Authors of winning tasks will be awarded prize money and invited to be co-authors on the resulting paper. Below, we detail our call for submissions. Feel free to join our Slack to message us with questions, find collaborators, and participate in contest-related discussions with other participants (code, ideas, findings, and related work sharing).

Prize information

2023/03/21 Update: The prize pool has been funded by Open Philanthropy

We will award up to $250,000 in total prize money for task submissions, distributed as follows:

1. Up to 1 Grand Prize of $100,000.
2. Up to 5 Second Prizes of $20,000 each.
3. Up to 10 Third Prizes of $5,000 each.

All prize decisions will be made by the organizers and anonymous reviewers, using the Prize Rubric below. Prize winners may nominate a non-profit to receive the prize money on their behalf. Some prizes may remain unawarded if there are not enough tasks that meet the eligibility for a prize tier, as detailed in the Prize Rubric.

Benchmark and Co-authorship: Authors of prize-winning submissions will be invited as co-authors on the paper written after the contest concludes. We will also offer co-authorship to authors of submissions that met our acceptability criteria but did not receive prizes, in the event that we receive more acceptable submissions than we can award with prizes. We will include all accepted submissions in our final benchmark, which we plan to release to the research community after the contest.

Timeline: The contest begins on June 27, 2022. We will host a first round of evaluations on submissions received on or before August 27, 2022 (Anywhere on Earth) and a second, final round of evaluations on submissions received on or before October 27, 2022 (Anywhere on Earth). After the first round, we will award eligible tasks with third prizes (up to 5) and second prizes (up to 2). To help improve first-round submissions, we will also return reviewer feedback and scaling law plots/results from our private, evaluation models. Submissions will be paused for two weeks at the end of the first round to allow any necessary improvements to be made. At the end of the second round, we will reward eligible tasks at all prize tiers, with the possibility of upgrading first-round submissions to higher prize tiers based on both rounds of submissions.

Prize Rubric

Here, we detail our submission evaluation rubric. The rubric will guide an anonymous panel of reviewers in judging submissions for prizes. A submission must meet all criteria in the "Grand Prize" column to win the grand prize. Likewise, a submission must meet all criteria in the "Accepted Task" column to be accepted into our benchmark and for co-authorship on our paper. For second prizes, submissions must meet all "Accepted Task" criteria and some "Grand Prize" criteria. Third prizes must meet the "Accepted Task" criteria. We may receive more eligible submissions than we have prizes for a given tier. In this case, we will first break ties based on how many “Grand Prize” criteria are met and then by having reviewers make subjective rankings within tiers (e.g., more granular measures of how much various criteria are met or the relative difficulty or importance of each criterion met). We will consider inverse scaling trends on publicly-available models like GPT-3, as well as held-out, private models for which we will run evaluation.

Answering the below, optional questions in our submission form (in the free-form response) will make your task stand out more:

• Does inverse scaling persist even if the model is conditioned with few-shot examples to behave correctly? If providing enough few-shot examples eliminates inverse scaling, how many examples are required for that?
• Does inverse scaling persist even after fine-tuning on the task? Are there good reasons to think it would persist after fine-tuning?
• Does inverse scaling persist for InstructGPT models trained with Reinforcement Learning from Human Feedback (RLHF)? To test this, you can use the same code as that for GPT-3 evaluation. We may also evaluate submissions on private RLHF models of various sizes from Anthropic [Bai et al. 2022].

We reserve the right to update the prize tier standards or criteria, e.g., between rounds if we observe submissions gaming them in some way.

Evaluation Eligibility: To be eligible for official review, a task submission must:

1. Include a plot of loss vs. model size across ada, babbage, curie, and davinci GPT-3 models, using the provided code for GPT-3 evaluation. The plot must not show a standard scaling law. A very bumpy trend is okay for submission; we expect to observe cleaner scaling laws with our held-out evaluation models, where we observe clear scaling trends.
2. Meet the formatting requirements described in the Submission Guidelines.
   • This requirement should already be satisfied if you are able to successfully run the evaluation code.
3. Include a coherent description of the task.

Models

This contest uses pretrained autoregressive language models such as GPT-3. We offer Google colab notebooks for evaluating inverse scaling with the GPT-3, OPT, and GPT-2 model series when developing a task. However, to avoid overfitting to publicly available models, we use private models to run the evaluations for awarding prizes. Currently, we are using the series of pretrained language models (without additional finetuning) from Anthropic [Bai et al. 2022]. We are in discussions with other organizations to use their models, which may be added later on to strengthen the evaluation.

Reviewers

Prize decisions will be made by an anonymous panel of reviewers. Reviewers will be selected by the contest organizers and may include some organizers. Reviewers will have ML and NLP experience relevant to inverse scaling. The panel may contain some competition organizers. Reviewers will not be allowed to make submissions to the contest.

Submission guidelines

1. Each task submission should be a language modeling test set (in the style of BIG-Bench) of inputs with corresponding answers, which will be evaluated according to one of four evaluation metrics (detailed later).
2. This prize is to incentivize original work, so submissions should find a new phenomenon for which inverse scaling has not been previously documented.
   1. If a task has already shown inverse scaling in prior work (even if the original authors did not identify it as such) then it is ineligible for the contest.
   2. If an existing task has not been subjected to any kind of scaling analysis, then it is likely eligible for the contest.
   3. If you would like to check whether an existing task is eligible, message us on our Slack or email us at inverse.scaling@gmail.com with [PRIOR WORK] in the subject line and a link to where the task has previously been published.
3. Data must be formatted as a .zip containing .csv files.
   • The zip should be called <task_name>.zip, where task_name is the name you provide for your submission in the form (e.g. lambada.zip).
   • The file should be called <task_name>.csv (e.g. lambada.csv).
     • If you have multiple parts to the same task, add -PART<i> to each (e.g. lambada-PART1.csv, lambada-PART2.csv).
     • If you have control experiments, add these as <task_name>-CONTROL<i>.csv (e.g. lambada-CONTROL1.csv).
   • The .csv files will be read using the pandas package, using the default arguments.
   • Specific formats are given below in the Evaluation metrics section.
4. Examples will be given as a prompt to an autoregressive language model.
   • I.e., either zero-shot or few-shot prompts (prompts containing a few examples). Few-shot examples must demonstrate the correct behavior on the task.
5. Tasks must contain at least 300 examples
   • We strongly recommend aiming for on the order of 1000 examples so that inverse scaling trends are clearer.
   • Submissions with unclear scaling trends and close to the minimum number of examples are unlikely to win a prize.
6. In the submission form, you will be asked to add:
   1. Evaluation metric used
      • The metric should be one of:
        1. Classification loss in a multiple-choice format (classification).
        2. Loss on a sequence at the end of the prompt (sequence_prob).
        3. Difference in logodds between two possible responses (logodds).
        4. Absolute difference in logodds between two possible responses (absolute_logodds).
   2. Authors
   3. Task name
   4. Description of intended task
      • What is the task aiming to test?
      • Remember to explain what good behavior on the task looks like, rather than just how the language model will do it wrong.
   5. Why is the task important?
      • How is the task important to safe and responsible use of LMs?
      • Does inverse scaling on the task suggest any fundamental insights about model behavior and model failures?
   6. Why do you expect to see inverse scaling?
      • What led you to suspect that your task will inverse scale?
   7. Data generation procedure
      • How did you construct the dataset? What resources did you use?
      • E.g., based on a template given pre-set vocabulary, programmatically generated based on some math phenomenon, etc.
      • If you used any code to generate the dataset, we ask that you include the code.
        • This is extremely helpful for assessing reproducibility of the results.
   8. Expertise required for human annotators to verify the task labels
      • E.g., knowledge of linguistics, fluent in specific languages.
   9. A plot of GPT-3 performance on the task, produced using this colab.
      • The plot must be generated using the unmodified colab to demonstrate that your dataset can be run through our code.
7. We will verify or validate task labels for correctness using contractors.
   • See this colab to test an interface similar to what we will use for task validation.
8. If you submit a very large dataset, we may uniformly subsample up to 10,000 examples from it for evaluation on private models.

Evaluation metrics

Here, we include more detail on the types of evaluation metrics that an eligible task submission can use. For examples of datasets correctly formatted for each metric, see this sheet. Please message us on our Slack or email us at inverse.scaling@gmail.com with [METRIC] in the subject line if you think it is impossible to frame your task as any of these evaluation metrics and that some other metric would demonstrate inverse scaling.

NOTE: We’ve identified a potential source of spurious inverse scaling with the logodds and absolute_logodds metrics. In particular, putting unrelated tasks for prompt and other_prompt can show inverse scaling, depending on the scaling behavior of the two tasks. A special case of this is adding irrelevant information to one of the prompts (thanks to Jeffrey Shen for pointing this out). We recommend framing your task as a classification or sequence probability task if possible unless you have an especially clear and considered justification for these metrics.

Classification loss (classification)

1. Use this metric for a standard classification task, e.g., when testing how well a model can choose the correct response from a list of possible responses.
   • We'll look at both classification loss and classification accuracy.
     • (Use classification_acc in the colabs to plot accuracy.)
   • If you expect inverse scaling to only appear on one of these metrics, explain why in the submission form.
2. The columns of the .csv are:
   1. prompt, a string.
   2. classes, an ordered list (using valid Python syntax) of the possible answer tokens.
      • NOTE: because of the way language model tokenizers typically process text, the class tokens should almost always start with a space and a capital letter, e.g. " Yes" and " No" rather than "yes" and "no".
      • Each class may be multiple tokens long, in which case we will use the logprob of the full token sequence as the logit for computing the classification loss. We strongly prefer for you to design each class to be the same number of tokens long, so that the sequence length of class labels does not affect the results (longer sequences tend to have smaller probabilities). Furthermore, we strongly encourage you to design class labels that are just a single token long. To do so, you can turn a task with a several possible sequence completions into an explicit, multiple choice task, where the classes are e.g. " A", " B", " C", and " D". This task format can guide the language model more clearly to put a high probability on valid completions.
      • As of Round 2, each example in the dataset can have a different number of possible classes.
   3. answer_index, an integer specifying which of the class tokens is the correct answer (using zero-indexing).

Completion sequence loss (sequence_prob)

1. Use this metric when you are testing how well the model predicts the correct conclusion or completion to a prompt.
   • This type of loss is used by LAMBADA, which uses as the target a single word. LAMBADA uses sequence probabilities because each example has only one correct completion, which is always a single word but may be multiple tokens long after tokenization.
   • Where possible, prefer using the classification loss (described above) instead of this metric, as sequence probabilities are often hard to interpret; there may be multiple correct completions for a given prompt that compete for probability mass (see Holtzmann et al. 2021 for discussion). For this reason, tasks that use classification loss will be more likely to pass our contractor task validation process.
2. The columns of the .csv are:
   1. prompt, a string up until the target word.
   2. completion, the target sequence (can be more than one token).

Signed difference in log-odds between examples (logodds)

1. This loss metric measures how much the model changes its probabilities between the two prompts.
   • In particular, the model can get better overall while still showing inverse scaling on this metric if the difference in log-odds between the two prompts grows.
   • Use this metric when testing how much changing the prompt influences the relative probabilities of the two class tokens in some undesirable way (e.g., to show bias against women). We will evaluate the average, signed (not absolute) difference between two prompts.
     • The first prompt, called prompt, should be neutral or favor the correct/desirable answer (e.g., a question about men, if expecting a male-favored bias).
     • The second prompt, called other_prompt, should be neutral or favor an incorrect/undesirable answer (e.g., a question about women, if expecting a male-favored bias).
2. As noted above, this metric requires a strong justification to use as it is prone to spurious inverse scaling.
   • If you feel that your task really requires the logodds metric, then we would encourage you to perform control experiments demonstrating that the inverse scaling is legitimate. This could look like replacing whatever you expect to cause the inverse scaling with neutral phrasing.
   • Please include descriptions and plots of these experiments in the free-form PDF response and upload the control data as part of your submission with names <task_name>-CONTROL<X>.csv where X is the index of your control experiment (e.g. lambada-CONTROL1.csv).
3. The columns of the .csv are:
   1. prompt, a string.
      • prompt should be neutral or be expected to favor the correct answer.
   2. other_prompt, a string.
      • other_prompt should be neutral or be expected to favor the incorrect answer.
   3. classes, an ordered list (using valid Python syntax) of the two possible answer tokens.
      • NOTE: because of the way tokenizers process text, the class tokens should almost always start with a space and a capital letter, e.g. " Yes" and " No" rather than "yes" and "no".
   4. answer_index, an integer specifying the index of the "correct" answer in the classes list (using zero-indexing).

Absolute difference in log-odds between examples (absolute_logodds)

1. This loss metric is identical to the signed difference (logodds) except that the absolute value of each log-odds difference is taken before averaging across examples.
   • This is useful when you expect there to be some difference between the prompts but are not sure of the direction.
   • An example is slightly varying the prompts in ways that should be irrelevant to the prediction: see True Few-shot Learning with Language Models.
   • Please use the logodds metric instead if you expect the difference between prompts to go in a particular direction (e.g., a consistent, male-favored bias).
2. As noted above, this metric requires a strong justification to use as it is prone to spurious inverse scaling.
   • If you feel that your task really requires the absolute_logodds metric, then we would encourage you to perform control experiments demonstrating that the inverse scaling is legitimate. This could look like replacing whatever you expect to cause the inverse scaling with neutral phrasing.
   • Please include descriptions and plots of these experiments in the free-form PDF response and upload the control data as part of your submission with names <task_name>-CONTROL<X>.csv where X is the index of your control experiment (e.g. lambada-CONTROL1.csv).
3. The columns of the .csv are:
   1. prompt, a string.
      • prompt should be one version of the prompt.
   2. other_prompt, a string.
      • other_prompt should be another version of the prompt, which should produce some difference in log-odds.
   3. classes, an ordered list (using valid Python syntax) of the two possible answer tokens.
      • NOTE: because of the way tokenizers process text, the class tokens should almost always start with a space and a capital letter, e.g. " Yes" and " No" rather than "yes" and "no".
   4. answer_index, an integer specifying the index of the "correct" answer in the classes list (using zero-indexing).
      • Note that this index does not matter for the absolute difference metric, but is needed for compatibility.

Submission instructions

1. Create a .csv file with your dataset examples following the guidelines
   1. If you have created your dataset as a Google Sheet, click:
      • File > Download > Comma Separated Values (.csv)
   2. See this sheet for examples.
2. Upload it to the GPT-3 colab and run it.
   • If you don’t have access to the OpenAI API, follow the instructions in the colab.
   • Based on our calculations, we expect the free credits provided by OpenAI to be enough to evaluate most tasks at least once across all model sizes.
     • During development, we recommend the OPT colab. In general, we recommend finding tasks that show inverse scaling with both GPT-3 and OPT, as such tasks are more likely to show inverse scaling during our evaluation procedure on held-out private models.
       • We also have a GPT-2 colab, but in our experience, the OPT colab demonstrates clearer scaling trends that generalize better to different and larger models.
     • Please fill in this form if you are unable to participate due to a lack of funding for credits.
       • If we are unable to support you directly, there is also this form from OpenAI for obtaining research credits directly from OpenAI.
3. Download the scaling plot produced.
4. Fill in the submission form.

NOTE: Some participants have mentioned that they are seeing consistent anomalies in scaling trends with certain OpenAI models (e.g., babbage). OpenAI isn't transparent about the details of their API models, so it's possible that the model is different in a way that breaks clean scaling law trends (e.g., trained the model for longer or on a different dataset). If you're seeing anomalous behavior from GPT-3 models, you should feel free to submit the tasks with anomalies in the scaling trends anyways (as long as the trend isn't a standard scaling law); we expect to get clearer signal about scaling laws with our held-out models from Anthropic, where we observe clean scaling laws and have more model sizes. We also recommend testing your submissions on the open-source OPT models (e.g. using our colabs), where we also see clear scaling laws and know the exact model sizes and training details; feel free to include the OPT plots in your submission as well, especially if you see clearer scaling trends.

FAQ

1. How does submitting to this contest interact with our own publication of the dataset/results?
   • We impose no restrictions on publication of submissions – in fact, we encourage it!
   • In addition, we will have performed a manual verification of the task labels and a scaling laws analysis on your data which may reduce the work required to publish.
2. Can people without machine learning expertise participate?
   • Yes! Our resources (such as the GPT-3 colab for testing your task) are designed to be useable with no coding – just provide the data.
   • Here are some example Google sheets for each of the evaluation metrics: Inverse Scaling Dataset Examples.
   • Simply click File > Download > Comma Separated Values (.csv) to get a .csv file that can be submitted.
3. Is there a restriction on what language the submission can be in? Is any arbitrary unicode script supported?
   • All languages are accepted so long as you can help us find contractors who are able to perform the task for label verification.
4. Relatedly, if I have a task that I’ve translated into 10 languages, is that one multilingual task or 10 monolingual tasks?
   • Translations of a task count as the same task – please only submit each example in one language.
5. How many tasks can I submit?
   • Contestants can win an award for each distinct submission; there's no limit to the number of awards one can receive. That said, a single submission that covers a broad phenomenon is more likely to win a higher tier prize due to our "task importance" criterion, so contestants may wish to group tasks together when possible.
   • If we receive too many submissions from one contestant, we may ask the contestant to choose a subset of their submissions for evaluation, especially if the submissions seem low-quality.
   • If you think that using multiple distinct datasets to test a single underlying phenomenon is called for, include them as separate .csv files in the .zip file of a single submission.
   • If you want to make changes to your submission please edit your original submission rather than resubmitting.
6. I’ve checked all labels in the submitted data, can I just report that instead of having your team verify it?
   • For consistency, all tasks will have to pass our review, even if they have undergone other validation.
7. Why do we require task submissions to have at least 300 examples?
   • 300 examples was the minimum number of examples for which we found clean, standard scaling trends on LAMBADA. LAMBADA is a next word prediction (language-modeling-like) task where we expect to observe clear, standard scaling laws, so we expect 300 examples to be a rough minimum number of examples to observe (inverse) scaling laws. However, we strongly encourage aiming for roughly 1000 examples, as trends are quite unclear for some tasks with only 300 examples.
8. Should prompts include instructions?
   • For tasks where you expect that a typical NLP practitioner would try instruction prompting, we'd encourage you to formulate your task in this as well; we will consider your task submission more important and realistic if it shows inverse scaling even with clear instructions in the example prompts. Please make sure to format the instructions in a way that you should reasonably expect a pretrained LM (i.e., GPT-3 or OPT, rather than InstructGPT) should be able to understand.
9. Should prompts include few-shot example sequences?
   • We encourage you to look for tasks that show inverse scaling even with in-context few-shot learning (as in the GPT-3 paper). For many tasks where a model does poorly, NLP practitioners would try few-shot learning. We will consider your task submission more important and realistic if it shows inverse scaling even with correctly-labeled few-shot examples of the task in the example prompts.
10. How will the prize money be distributed between co-authors of a task submission?
    • We will leave it up to the co-authors of submissions to decide how to distribute the prize money between co-authors.
11. How do I ask a question about the contest?
    • Please send a message in the #ask-organizers channel of the Inverse Scaling Slack for general or public questions that others may have. Please email us at inverse.scaling@gmail.com or message us individually on the Slack for private questions related to your specific submission, task, or situation.
12. What license should I use for my dataset?
    • Please use the CC-BY license.
13. Can I submit examples of misuse as a task? What about other tasks that explicitly prompt the model to produce harmful behavior?
    • We don't consider most cases of misuse or explicitly prompting for harmful behavior as surprising examples of inverse scaling. We expect that explicitly prompting/asking an LM to generate e.g. hate speech or propaganda will work more effectively with larger models, so we do not consider such behavior surprising. That said, we are still interested novel misuse examples, but they will need to be surprising in order to be accepted.
14. Can I submit examples of bias, toxicity, or misinformation as a task?
    • In many of the above cases, inverse scaling has been shown to some extent in prior work. For example, [Solaiman & Dennison 2021] showed that larger GPT-3 models generate more toxic text, [BIG-Bench] showed that larger models show more bias related to gender, race, ethnicity, and nationality, and [TruthfulQA] suggests that larger models generate more plausible misinformation. As a result, we will apply more scrutiny for novelty on submissions on or other topics covered by prior work. That said, we are still interested in submissions related to these topics, if submissions effectively argue they have found a different phenomenon (e.g., a novel kind of bias) or that prior work did not sufficiently show inverse scaling for some behavior.
15. Can I submit a task that demonstrates positive scaling on the incorrect answer?
    • As of Round 2, this will be considered.
    • We still prefer a demonstration of inverse scaling. However, if there is exactly one incorrect option in classification or exactly one completion that can be considered 'incorrect' in sequence_prob, then we will consider tasks that show positive scaling on that answer.

Tips

• Again, feel free to join our Slack to message us with questions, find collaborators, and participate in contest-related discussions with other participants (code, ideas, findings, and related work sharing). Joining our Slack isn't necessary, but we created it to facilitate relevant discussions, content-sharing, and collaborations.
• While we accept narrow tasks with as few as 300 examples, we believe that large, diverse datasets are more likely to do well on our criteria (e.g., to show consistent, clean inverse scaling on held-out models and be of broader importance). In particular, inverse scaling results may be sensitive to the particular prompt used [9], in which case the results may not generalize to held-out models; phrasing examples in various ways should reduce this effect.
• It may be helpful to quickly iterate on your task submission by hand-writing a small number of examples first or by using a template or code to construct a larger dataset automatically. After finding something that scales inversely, you may improve the number and diversity of examples for your final submission (for reasons stated above).
• We encourage looking at the error bars provided by our colabs to understand the variance in performance across examples, to get a sense of how clean vs. noisy the trend may be on our private, evaluation models.
• We copied the OPT model family (up to OPT-66B) to our Huggingface Hub organization (inverse-scaling/opt-...) so you can run evaluations of your Huggingface datasets with the new Huggingface model-evaluator. Running the model evaluation will create a PR we can then merge into our copied model repo (which is why we keep a copy of the original model repo). For the required dataset format, see the winning Round 1 datasets on our Huggingface organization page. (There is also a small commandline tool to convert your .csv files to .jsonl files for your Huggingface dataset.) NOTE: Due to the way the model-evaluator works, your datasets will be publicly visible.

Links

Here are all the relevant links that appear in these instructions.

• Join our Slack
• Submission form
• Task Examples Google Sheets
• GPT-3 Colab
• OPT Colab
• GPT-2 Colab
• Validation Example Colab

About Us

2023/03/21 Update: The prize pool has been funded by Open Philanthropy.

2022/12/12 Update: The prize pool is no longer provided by FTX Future Fund.

The Inverse Scaling Prize is organized by a group of researchers on behalf of the Fund for Alignment Research (FAR), including Ian McKenzie, Alexander Lyzhov, Michael Pieler, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. Additionally, Sam Bowman and Ethan Perez are affiliated with Anthropic; Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman are affiliated with New York University. The prize pool is provided by Open Philanthropy.

Prize Money Eligibility

To minimize conflict of interest, we will not be able to give prize money to people at Anthropic, New York University, FAR AI, or the Future Fund. Authors of task submissions who come from these institutions will still be eligible to win an award without prize money, as long as they are not organizers or anonymous reviewers. For legal reasons, we will also not be able to award prize money to individuals sanctioned by the U.S. It is possible that we will add more evaluation model series from other institutions. In this event, authors of task submissions from those institutions will have their submissions evaluated on model series from other participating institutions (e.g. the Anthropic models) and will still be eligible for prizes.

Prize Terms and Conditions

These terms apply to the second round, and don't necessarily apply to the first round. If these criteria might prevent you from making a submission, please let us know over email or Slack.

• Employees or current contractors of FTX Philanthropy and contest organizers are not eligible to win prizes.
• Entrants and Winners must be over the age of 18.
• By entering the contest, entrants agree to the Terms & Conditions.
• All taxes are the responsibility of the winners.
• The legality of accepting the prize in his or her country is the responsibility of the winners. Sponsor may confirm the legality of sending prize money to winners who are residents of countries outside of the United States.
• Winners will be notified by email after the contest deadline.
• Winners grant to Sponsor the right to use their name and likeness for any purpose arising out of or related to the contest. Winners also grant to Sponsor a non-exclusive royalty-free license to reprint, publish and/or use the entry for any purpose arising out of related to the contest including linking to or re-publishing the work.
• Entrants warrant that they are eligible to receive the prize money from any relevant employer or from a contract standpoint.
• Entrants agree that FTX Philanthropy shall not be liable to entrants for any type of damages that arise out of or are related to the contest and/or the prizes.
• By submitting an entry, entrant represents and warrants that, consistent with the terms of the Terms and Conditions: (a) the entry is entrant’s original work; (b) entrant owns any copyright applicable to the entry; (c) the entry does not violate, in whole or in part, any existing copyright, trademark, patent or any other intellectual property right of any other person, organization or entity; (d) entrant has confirmed and is unaware of any contractual obligations entrant has which may be inconsistent with these Terms and Conditions and the rights entrant is required to have in the entry, including but not limited to any prohibitions, obligations or limitations arising from any current or former employment arrangement entrant may have; (e) entrant is not disclosing the confidential, trade secret or proprietary information of any other person or entity, including any obligation entrant may have in connection arising from any current or former employment, without authorization or a license; and (f) entrant has full power and all legal rights to submit an entry in full compliance with these Terms and Conditions.

Citations

To cite the contest, you may use the following bibtex citation:

@misc{mckenzie2022inverse,
    title={The Inverse Scaling Prize},
    url={https://github.com/inverse-scaling/prize},
    author={McKenzie, Ian and Lyzhov, Alexander and Parrish, Alicia and Prabhu, Ameya and Mueller, Aaron and Kim, Najoung and Bowman, Sam and Perez, Ethan},
    year={2022}
}

To cite the Round 1 winners, you may use:

@misc{mckenzie2022round1,
    title={Inverse Scaling Prize: First Round Winners},
    url={https://irmckenzie.co.uk/round1},
    author={McKenzie, Ian and Lyzhov, Alexander and Parrish, Alicia and Prabhu, Ameya and Mueller, Aaron and Kim, Najoung and Bowman, Sam and Perez, Ethan},
    year={2022}
}

To cite the Round 2 winners, you may use:

@misc{mckenzie2022round2,
    title={Inverse Scaling Prize: Second Round Winners},
    url={https://irmckenzie.co.uk/round2},
    author={McKenzie, Ian and Lyzhov, Alexander and Parrish, Alicia and Prabhu, Ameya and Mueller, Aaron and Kim, Najoung and Bowman, Sam and Perez, Ethan},
    year={2023}
}

---

• [1] Brown et al. Language Models Are Few-Shot Learners. NeurIPS 2020.
• [2] Chowdhery et al. PaLM: Scaling Language Modeling with Pathways. arXiv 2022.
• [3] Liang et al. Towards Understanding and Mitigating Social Biases in Language Models. ICML 2021.
• [4] Lin, Hilton, and Evans. TruthfulQA: Measuring How Models Mimic Human Falsehoods. ACL 2022.
• [5] Kaplan et al. Scaling Laws for Neural Language Models. arXiv 2020.
• [6] Hoffmann et al. Training Compute-Optimal Large Language Models. arXiv 2022.
• [7] Hernandez et al. Scaling Laws for Transfer. arXiv 2021.
• [8] Parrish et al. BBQ: A Hand-Built Bias Benchmark for Question Answering. ACL 2022 Findings.
• [9] Perez et al. True Few-Shot Learning with Language Models. NeurIPS 2021.
• [10] Srivastava et al. Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models. arXiv 2021.