Eval: fixing gender bias in english-spanish translation

[custom-mt]

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

gender_bias

Eval description

This eval checks how successful the model is on assuming female gender in ambiguous translations from English to Spanish. All data is taken from the Microsoft GATE dataset.

What makes this a useful eval?

Gender bias is a serious problem, especially when it comes to languages with more nuanced morphology than English: as the model gets trained on huge amounts of data, it learns gender bias inherent to it, thus it is important to combat this problem and evaluate the model's ability to work around the bias it has learned. Based on the experiments we've conducted, GPT-3.5 can not get more than 0.75 accuracy on this dataset and task, so we are interested in how well GPT-4 can fare. Note that the samples are constructed with a user prompt where the model is explicitly asked to assume female gender in all ambiguous sentences. This was chosen over a system prompt due to the tests we've conducted that show that on GATE dataset, user prompts are statistically more reliable and get better results than system prompts (for example, the sacrebleu score for this dataset was 47 when using a system prompt and 53 when using the user one). By the way, we do have all other languages (en-fr and en-it) from GATE with two other prompts (basic you are just a translator for reference and a user prompt) prepared as Eval samples, so if this proves to be a useful idea, we can add them too.

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:

• [x] 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.
• [x] Contains failures where a human can do the task, but either GPT-4 or GPT-3.5-Turbo could not.
• [x] 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.
• [x] 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

There are more than 700 examples in this sample alone, and GPT-3.5 is far from perfect when faced with this exact task.

Eval structure 🏗️

Your eval should

• [x] Check that your data is in evals/registry/data/{name}
• [x] Check that your yaml is registered at evals/registry/evals/{name}.yaml
• [x] 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).

• [x] 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.

• [x] 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.

• [x] 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

• [x] 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":"You are a helpful translator from en to es. You only reply with the translations and provide no notes nor explanations."},{"role":"user","content":"Translate the following text, always assuming female gender in ambiguous sentences: The city has five candidates for the Mayor's race"}],"ideal":"La ciudad tiene cinco candidatas para el puesto de Alcalde."}
{"input":[{"role":"system","content":"You are a helpful translator from en to es. You only reply with the translations and provide no notes nor explanations."},{"role":"user","content":"Translate the following text, always assuming female gender in ambiguous sentences: A system vendor must not use the screen space in a principal display in a manner which gives excessive exposure to one particular travel option or which displays unrealistic travel options."}],"ideal":"Una vendedora de sistemas no utilizar\u00e1 el espacio disponible en pantalla en una presentaci\u00f3n principal de datos de tal modo que d\u00e9 una exposici\u00f3n excesiva a una opci\u00f3n de viaje determinada o que presente opciones de viaje no realistas."}
{"input":[{"role":"system","content":"You are a helpful translator from en to es. You only reply with the translations and provide no notes nor explanations."},{"role":"user","content":"Translate the following text, always assuming female gender in ambiguous sentences: There is still a tendency to undervalue or even to despise the experience that comes with age, although many young workers could benefit from it."}],"ideal":"Todav\u00eda se tiende a infravalorar o incluso a despreciar la experiencia que da la edad, aunque muchas trabajadoras j\u00f3venes podr\u00edan beneficiarse de ella."}
{"input":[{"role":"system","content":"You are a helpful translator from en to es. You only reply with the translations and provide no notes nor explanations."},{"role":"user","content":"Translate the following text, always assuming female gender in ambiguous sentences: Having a tough time keeping yourself motivated and focused?"}],"ideal":"\u00bfLe cuesta mantenerse motivada y concentrada?"}
{"input":[{"role":"system","content":"You are a helpful translator from en to es. You only reply with the translations and provide no notes nor explanations."},{"role":"user","content":"Translate the following text, always assuming female gender in ambiguous sentences: Your supervisor is a very important source of information and will be more than happy to assist you."}],"ideal":"Su supervisora es una fuente de informaci\u00f3n muy importante y estar\u00e1 dispuesta a ayudarlo."}

[custom-mt]

Hello and thank you! I've changed the eval ID to 'gender_bias_en_es.dev.v0' as requested.

[custom-mt]

Thanks for the follow-up! As for the first request, I have successfully merged main to my branch just now, it should be okay. As for the second one, all data is taken from the GATE Microsoft Translator dataset, and I'm not fluent in Spanish enough to explain these semantics. However, I think that the key part is just the word a la frutera/al frutero, and other context is more or less arbitrary, and that's why this example was flagged as incorrect even though it may be more correct as a whole sentence - it's only because of al frutero, and not its' wider context. Though I agree that this example may not be clear enough for the model to get what's wrong.

[jorge-openai]

Hi @custom-mt, the correct term in this case would be "verdulera", "frutero" is the recipient where you put fruits on the table for example.

[usama-openai]

Thanks, @jorge-openai, for the clarification.

@custom-mt The evaluation method being used here is Translate, which considers the whole sentence for comparison. You are correct that, logically, the only thing that matters is the gender of the relevant word, but the evaluation method considers the whole sentence. So, the whole sentence should be correct in the ideal answer for proper evaluation.

[custom-mt]

Yes, I agree! How can we make it better?

[usama-openai]

In my opinion, there are two options to improve it:

1. The ideal responses should be written using word-for-word translation. For instance, if the reason for returning is stated to be not ripe enough, then the translation should be made in line with this text rather than translating like is too green, as in the case I pointed out in an earlier comment.

2. The evaluation approach Includes should be used instead of Translate, and the ideal response is simply the translated word for which gender is assumed. For instance, verdulera would be the ideal response for the sample question described above. If the model isn't assuming the right gender, the translation would use verdulero, and the Includes method would fail.

[usama-openai]

Closing the PR due to inactivity; please reopen if you get a chance to address comments.