Specialized Translation Model for Social Work

Our Mockup

Guided by humanist implementation and data justice principles, built to benefit communities and front line social workers.

Prompt Engineering Examples

These prompts will guide the model's responses. Copy and paste into the model interface to test them out: "Try the Model"> scroll down to "Additional Inputs"> paste into "System Prompt"

Example Prompt 1
Prompt

You are a social work assistant. Translate the following message using strengths-based, plain language. Avoid deficit framing.

Why it works

Explicitly naming the framework upfront ("strengths-based, plain language") shapes the model's register before it generates a response.

Example Prompt 2
Prompt

Translate the following with fidelity, add an annotation for concepts that may need cultural context for an American social worker to understand

Why it works

This prompt ensures that the model translates the text accurately while also providing additional context that may be necessary for a different cultural audience.

Example Prompt 3
Prompt

Contribute an Idea to the Community.

Submit your own prompt idea

Try it in the Model

Interact with tiny-anya directly below. You can also open it in a new tab for a full-screen experience.

Open in full screen Opens on Hugging Face Spaces ↗
huggingface.co/spaces/CohereLabs/tiny-aya

Evaluation Examples

Evaluation will include AND go beyond standard metrics, and it's imperitive that practictioners and community members are involved in development of evaluation.

Potential Evaluation Dimension 1

Meaning Preservation

For each translation, does the model maintain the intended meaning across different languages and contexts? Can be Y/N, or a scale (1-5, for example)

Potential Evaluation Dimension 2

Cultural Responsiveness

For each translation, does the model adapt its output appropriately to the specified cultural and linguistic context? Can be Y/N, or a scale (1-5, for example)

Potential Evaluation Dimension 3

Bias & Representation

For each translation, is there an identifiable bias in the output?

Potential Evaluation Dimension 4

Practitioner Notes of Interest and Concern

What observations or concerns did practitioners or peers raise when reviewing model outputs? Can be free text, that is then categorized and analyzed so it can be used to inform future improvements.

Contribute a Community Evaluation Criterion

Add to our list of ideas below.

Submit an idea

Community Ideas

Have a prompt idea or evaluation criterion to suggest? Add it to the shared Notion board below — everyone can see and build on each other's contributions.