Add Grid Position Tracking eval

[adamowada]

Thank you for contributing an eval! ♥️

🚨 Please make sure your PR follows these guidelines, failure to follow the guidelines below will result in the PR being closed automatically. Note that even if the criteria are met, that does not guarantee the PR will be merged nor GPT-4 access granted. 🚨

PLEASE READ THIS:

In order for a PR to be merged, it must fail on GPT-4. We are aware that right now, users do not have access, so you will not be able to tell if the eval fails or not. Please run your eval with GPT-3.5-Turbo, but keep in mind as we run the eval, if GPT-4 gets higher than 90% on the eval, we will likely reject since GPT-4 is already capable of completing the task.

We plan to roll out a way for users submitting evals to see the eval performance on GPT-4 soon. Stay tuned! Until then, you will not be able to see the eval performance on GPT-4. Starting April 10, the minimum eval count is 15 samples, we hope this makes it easier to create and contribute evals.

Eval details 📑

Eval name

Grid Position Tracking

Eval description

Tests a model's proficiency in maintaining an accurate count of steps taken and determining the final coordinate position after a series of steps on a grid. Each test starts at the origin (0, 0), with each step moving the position 1 space in one of four directions: up, down, left, or right. The sequence of steps is randomly generated, and the eval includes 100 tests, ranging from 2 to 101 steps.

What makes this a useful eval?

It serves as a straightforward method for examining a model's memory and spatial reasoning capabilities in a simulated physical environment. By streamlining the movement and employing incremental testing, this eval facilitates the identification of a model's point of failure. Conducting repeated evaluations enables a better understanding of the statistical likelihood of model failure which supports precise A/B testing. For example, after 1,000 runs, it could be determined that Model A is more likely to fail than not after 35 steps, while Model B tends to fail after 37 steps.

This eval is useful for LLMs for several reasons:

1. Math / logical / physical reasoning: It specifically tests an LLM's ability to perform mathematical calculations, logical deductions, and spatial reasoning, simulating a physical environment in a grid-based setting.

2. Real-world use case: Spatial reasoning and memory capabilities are crucial for LLMs deployed in various applications such as robotics, navigation systems, and virtual agents. This eval provides a foundation for understanding a model's performance in these real-world scenarios.

3. Other foundational capability: The eval helps to assess a model's working memory, which is a key aspect of its cognitive abilities. Working memory is vital for LLMs as it influences their capacity to process and retain information in tasks requiring complex reasoning.

Criteria for a good eval ✅

Below are some of the criteria we look for in a good eval. In general, we are seeking cases where the model does not do a good job despite being capable of generating a good response (note that there are some things large language models cannot do, so those would not make good evals).

Your eval should be:

• Thematically consistent: The eval should be thematically consistent. We'd like to see a number of prompts all demonstrating some particular failure mode. For example, we can create an eval on cases where the model fails to reason about the physical world.
• Contains failures where a human can do the task, but either GPT-4 or GPT-3.5-Turbo could not.
• Includes good signal around what is the right behavior. This means either a correct answer for Basic evals or the Fact Model-graded eval, or an exhaustive rubric for evaluating answers for the Criteria Model-graded eval.
• Include at least 15 high quality examples.

If there is anything else that makes your eval worth including, please document it below.

Unique eval value

I've noticed a substantial improvement in model performance when tasked with tracking each move individually, as opposed to solely determining the final position. The underlying logic of this eval can be adapted to encompass step-by-step tracking and more intricate moves, instead of just increasing the number of steps involved.

Eval structure 🏗️

Your eval should

• Check that your data is in evals/registry/data/{name}
• Check that your yaml is registered at evals/registry/evals/{name}.yaml
• Ensure you have the right to use the data you submit via this eval

(For now, we will only be approving evals that use one of the existing eval classes. You may still write custom eval classes for your own cases, and we may consider merging them in the future.)

Final checklist 👀

Submission agreement

By contributing to Evals, you are agreeing to make your evaluation logic and data under the same MIT license as this repository. You must have adequate rights to upload any data used in an Eval. OpenAI reserves the right to use this data in future service improvements to our product. Contributions to OpenAI Evals will be subject to our usual Usage Policies (https://platform.openai.com/docs/usage-policies).

• I agree that my submission will be made available under an MIT license and complies with OpenAI's usage policies.

Email address validation

If your submission is accepted, we will be granting GPT-4 access to a limited number of contributors. Access will be given to the email address associated with the merged pull request.

• I acknowledge that GPT-4 access will only be granted, if applicable, to the email address used for my merged pull request.

Limited availability acknowledgement

We know that you might be excited to contribute to OpenAI's mission, help improve our models, and gain access to GPT-4. However, due to the requirements mentioned above and high volume of submissions, we will not be able to accept all submissions and thus not grant everyone who opens a PR GPT-4 access. We know this is disappointing, but we hope to set the right expectation before you open this PR.

• I understand that opening a PR, even if it meets the requirements above, does not guarantee the PR will be merged nor GPT-4 access granted.

Submit eval

• I have filled out all required fields in the evals PR form
• (Ignore if not submitting code) I have run pip install pre-commit; pre-commit install and have verified that black, isort, and autoflake are running when I commit and push

Failure to fill out all required fields will result in the PR being closed.

Eval JSON data

Since we are using Git LFS, we are asking eval submitters to add in as many Eval Samples (at least 5) from their contribution here:

View evals in JSON

Eval

{"input": [{"role": "system", "content": "Let's say you start at point (0, 0) on a grid. I will tell you a series of moves to make. Keep track of your position and the total number of steps taken. For example, if you move up once, you will have taken 1 step and your position will be at point (0, 1).\n\nYou will respond in the following format:\n\nAfter __ steps, the final position is (__, __)\n\nYou will fill in the blanks."}, {"role": "user", "content": "down, down"}], "ideal": "After 2 steps, the final position is (0, -2)"}
{"input": [{"role": "system", "content": "Let's say you start at point (0, 0) on a grid. I will tell you a series of moves to make. Keep track of your position and the total number of steps taken. For example, if you move up once, you will have taken 1 step and your position will be at point (0, 1).\n\nYou will respond in the following format:\n\nAfter __ steps, the final position is (__, __)\n\nYou will fill in the blanks."}, {"role": "user", "content": "left, right, left"}], "ideal": "After 3 steps, the final position is (-1, 0)"}
{"input": [{"role": "system", "content": "Let's say you start at point (0, 0) on a grid. I will tell you a series of moves to make. Keep track of your position and the total number of steps taken. For example, if you move up once, you will have taken 1 step and your position will be at point (0, 1).\n\nYou will respond in the following format:\n\nAfter __ steps, the final position is (__, __)\n\nYou will fill in the blanks."}, {"role": "user", "content": "right, right, right, right"}], "ideal": "After 4 steps, the final position is (4, 0)"}
{"input": [{"role": "system", "content": "Let's say you start at point (0, 0) on a grid. I will tell you a series of moves to make. Keep track of your position and the total number of steps taken. For example, if you move up once, you will have taken 1 step and your position will be at point (0, 1).\n\nYou will respond in the following format:\n\nAfter __ steps, the final position is (__, __)\n\nYou will fill in the blanks."}, {"role": "user", "content": "right, right, right, up, down"}], "ideal": "After 5 steps, the final position is (3, 0)"}
{"input": [{"role": "system", "content": "Let's say you start at point (0, 0) on a grid. I will tell you a series of moves to make. Keep track of your position and the total number of steps taken. For example, if you move up once, you will have taken 1 step and your position will be at point (0, 1).\n\nYou will respond in the following format:\n\nAfter __ steps, the final position is (__, __)\n\nYou will fill in the blanks."}, {"role": "user", "content": "right, down, right, left, down, up"}], "ideal": "After 6 steps, the final position is (1, -1)"}

[Ein-Tim]

Related PRs:

• 2D Grid Traversal Eval #763
• [Eval] Random walk on a 2D grid #462

[adamowada]

Related PRs:

• 2D Grid Traversal Eval #763
• [Eval] Random walk on a 2D grid #462

Thank you for pointing this out Tim!

[jorge-openai]

Thanks for opening this PR, counting is a well-known failure mode of the model due to a common underlying issue in LLMs. In its current form, this eval does not seem to expose any new gaps in our understanding of model performance. There is also an already merged PR with the same challenges of following the grid. We also know that this could be solved by giving the model a code interpreter.

If you're still interested in writing an eval, we've noticed that these criteria make good evals. If you have any particular use case in mind for the model, can you come up with an eval that has some of these attributes?

• Multi-step reasoning
• Domain or Application specific
• Open-Ended responses
• Complex instructions
• The eval seems obvious, but tricks the model in a novel way

I'm closing this PR, but feel free to open a new one with the above suggestions.