Goal: put your work into the world in a way that helps people and does not create problems for you.
Why this step matters¶
Open release is how a lone builder gets leverage. Publish your weights, tokenizer, cleaned data, and evaluation set, and people who care about the language will find you. Scholars will correct you. Contributors will bring data you could not have collected alone.
That only works if the release is honest and well documented.
What you do¶
1. Test safety in Sanskrit and Urdu, not in English¶
2. Red-team your own model¶
Try to make it fail. Write down what worked. For a model dealing with religious and philosophical text, pay particular attention to confident fabrication — a made-up verse presented as authentic is a serious failure.
3. Sort out your licences¶
You have four to consider:
Your data’s licence
Your base model’s licence
The licence of any teacher model whose outputs you used
The licence you release under
They stack, and the most restrictive one wins. Work this out before release, not after.
4. Write an honest model card¶
Include:
What it does well
What it does badly
What data it saw
How you evaluated it
The failure modes you found in step 2
A model card that lists real weaknesses is more trustworthy, not less. It also saves you from people discovering those weaknesses publicly and loudly.
5. Release properly¶
Weights, in both full and quantized forms
The tokenizer
Your evaluation results and the evaluation set itself
The training configuration
The data recipe, if you can
6. Make it trivially easy to run locally¶
Publish GGUF versions. Make sure it works with the common local tools. If someone has to fight your model for an hour, they will not.
7. Make it reproducible¶
Random seeds, exact library versions, hardware used, and the exact command you ran.
8. Release the evaluation set and the fertility numbers separately¶
Honestly, these may be more useful to the community than your model. Your model will be superseded within a year. A careful, native-checked evaluation set for classical Sanskrit will still be useful in ten.
9. Document the journey¶
If you write, record, or stream, do it. Not for vanity — for recruitment.
Sanskrit and Urdu both have passionate communities who are not AI engineers but who have exactly the domain knowledge you lack. A public build log is how they find you. Some of the most valuable contributions to a project like this come from people who cannot write a training loop but can tell you instantly that your model’s output is grammatically wrong.
You are ready to move on when¶
Someone you have never met has downloaded your model, run it, and told you something about it that you did not know.