# Grading rubric

## v1
| Criterion | Initial Consideration | Third Prize Tier | Second Prize Tier | Grand Prize Tier |
| --------- | --------------------- | ----------------- | ----------------- | ----------- |
| Slope | Doesn’t show regular scaling on public models. | Shows approximately monotonic inverse scaling on private models. | Shows non-decreasing inverse scaling on private models. | Shows clear, strictly monotonic inverse scaling on private models. |
| Task importance | The task is clearly specified. An argument is presented for why the task is important. | The task is clearly specified. There is a strong argument for why this task relates to some aspect of model behavior. | The task is clearly specified. There is a strong argument for why this task relates to some aspect of model behavior and an argument for why this aspect of model behavior is particularly important to the responsible use of LMs. | The task is clearly specified. It is made very clear how this task relates to model behavior and why this aspect of model behavior is crucial for the safe and responsible use of LMs (perhaps with an example of what could go wrong). |

## v2
| Aspect of evaluation | Description | Poor | Adequate | Perfect | |
| --------------------------------- | ------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | --- |
| Strength of inverse scaling | When we examine a scaling curve with task performance vs. model size, how clean is the inverse scaling trend? | It's questionable because the scaling curve looks very patchy and noisy. | Inverse scaling curve is approximately monotonic despite some noise and defects. | Inverse scaling is very clear. | |
| Universality of inverse scaling | When we compute task performance for another model family, do we keep seeing strong inverse scaling? | The effect is highly dependent on the model family and cannot be reproduced for others. | Inverse scaling transfers across public and private model families with some minor exceptions. | Inverse scaling is very clear when evaluating across different public and private model families. | |
| Task importance | How important is the task to safe use of LMs? | The arguments for task importance are weak. | The arguments for why the task is important to safe use of LMs are convincing. | The arguments for why the task is particularly important to safe use of LMs are convincing (perhaps with an example of what could go wrong). | |
| Novelty and unexpectedness | Is the observation of inverse scaling on the task novel? How unexpected is the effect? | Inverse scaling on the task is already well-known. | Inverse scaling on the task is a novel but not very surprising discovery. | Inverse scaling on the task is a novel and very surprising discovery, teaching us new important things about LMs. | |
| Rectifiability of inverse scaling | How hard it is to fix the phenomenon and make task performance scale up with model size? | It's very easy to fix inverse scaling by slightly changing the prompt format. | Inverse scaling almost always persists even after serious attempts to construct the prompts differently. | Inverse scaling is resistant to replacing the dataset with new prompts framing the same task in a different way, providing few-shot examples of correct behavior, fine-tuning on the task. | |
| Coverage of the task | Are the prompts fully representative of the described task? | Prompts only cover a particular and special subcategory of the described task. | Prompts cover the most important ways in which the task can be framed, even though many obvious subcategories of the task are omitted. | Prompts exhibit great diversity and cover all important ways in which the task can be framed. | |
| | | | | | |

## v3

| Aspect of evaluation | Description | Poor | Adequate | Perfect | |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --- |
| Strength of inverse scaling | When we examine a scaling curve with task performance vs. model size, how clean is the inverse scaling trend? | It's questionable because the scaling curve looks very patchy and noisy. | Inverse scaling curve is approximately monotonic despite some noise and defects. | Inverse scaling is very clear. | |
| Universality of inverse scaling | When we compute task performance for another model family, do we keep seeing strong inverse scaling? | The effect is highly dependent on the model family and cannot be reproduced for others. | Inverse scaling transfers across public and private model families with some minor exceptions. | Inverse scaling is very clear when evaluating across different public and private model families. | |
| Task importance | How important is the task to safe use of LMs? | The arguments for task importance are weak. | The arguments for why the task is important to safe use of LMs are convincing. | The arguments for why the task is particularly important to safe use of LMs are convincing (perhaps with an example of what could go wrong). | |
| Novelty and unexpectedness | Is the observation of inverse scaling on the task novel? How unexpected is the effect? | Inverse scaling on the task is already well-known. | Inverse scaling on the task is a novel but not very surprising discovery. | Inverse scaling on the task is a novel and very surprising discovery, teaching us new important things about LMs. | |
| Coverage of the task | Are the examples fully representative of the described task? | Examples only cover a special subcategory or phrasing of the task, and there's no inverse scaling on other ones. | The task includes diverse subcategories and phrasings. Reproducing the task based on its description would also yield inverse scaling. | Examples cover all important task subcategories and phrasings, suggesting it's hard to eliminate inverse scaling by changing how the task is framed. | |
| | | | | | |

## Evaluation of submissions
### Minimal requirements

Before being evaluated against our rubric, submissions must:
1. Include a plot of performance on the GPT-3 models.
* Use [this colab](https://colab.research.google.com/drive/1SGmUh0NbqSrRkWRUcmjg8BS5eU5qvJ0Y) to produce a plot.
2. Meets the formatting requirements.
* This should already be satisfied if you are able to evaluate the submission on public models.
3. Show inverse scaling on public models such as GPT-3; however, a flat line or a noisy plot with unclear direction of scaling are also acceptable.
4. Contain a coherent description of the task with an argument for why it is important for safe and responsible use of LMs.

### Rubric

This rubric presents important dimensions along which the submissions are scored. The scores on the rubric do not have a direct correspondence with prize tiers, but it will assist an anonymous panel of reviewers in judging submissions.

| ***Criterion*** | Description | Poor | Adequate | Perfect | |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --- |
| ***Inverse Scaling Strength*** | How straight and steep is the inverse scaling trend? | Shows flat, very bumpy, or standard scaling. | Shows approximately monotonic inverse scaling. | Shows a clear, strictly monotonic inverse scaling trend that can be well-fit with a power law (a line on a log-log plot). | |
| ***Inverse Scaling Generality*** | When we compute task performance for a different model family, do we still see inverse scaling? | No inverse scaling on private models. | Shows inverse scaling on some public and some private model families. | Shows inverse scaling across all public and private model families tested. | |
| ***Task Importance*** | Is the task important to the safe and responsible use of LMs, or for shedding light on where LMs fail? How strong are the arguments? | Weak. No users would be harmed, and the task does not shed light on where LMs fail. | Fairly convincing. Some LM users would be harmed by the discovered behavior, or the task sheds light on where LMs fail (e.g., [sensitivity to prompts](https://arxiv.org/abs/2105.11447)). | Very convincing. Significant implications for how LMs should be used (e.g., bias, toxicity, or misinformation -related behaviors) | |
| ***Novelty and Surprisingness*** | Is inverse scaling on the task novel and surprising? | Not novel or surprising | Novel but not surprising | Novel and surprising, teaching us something new and important about LMs. | |
| ***Task Coverage*** | Are the prompts fully representative of the described task? | Examples only cover a special subcategory of the task, and there's no inverse scaling on other ones. | Examples cover diverse task subcategories and phrasings. Reproducing the task based on its description would also yield inverse scaling. | Examples cover all important task subcategories and phrasings, suggesting robust inverse scaling. | |
| | | | | | |

The "task importance" criterion will carry notable weight especially for higher prize tiers. It's okay to submit a task that shows inverse scaling strongly but not universally or universally but not strongly.

Answering the below, optional questions in our submission form 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 models trained with [Reinforcement Learning from Human Feedback (RLHF)](https://openai.com/blog/instruction-following/)? To test this, you can use models from the [InstructGPT series](https://openai.com/blog/instruction-following/) in the [GPT-3 colab](https://colab.research.google.com/drive/1SGmUh0NbqSrRkWRUcmjg8BS5eU5qvJ0Y). We may also evaluate submissions on private RLHF models of various sizes from Anthropic [[Bai et al. 2022](https://arxiv.org/abs/2204.05862)].

We reserve the right to disqualify tasks for reasons not listed in this rubric. For example:
- The task labels fail human verification.
- The individual task examples are highly optimized based on how much inverse scaling they produce, which makes the data unrepresentative of the task.