The first product in Thembi is a review system

The first product in Thembi is not a trained language model. It is a review system that helps fluent Xitsonga speakers decide which language examples are trustworthy enough to train or evaluate one.

That order is deliberate. A larger dataset or a completed fine-tuning run would create technical output, but neither would answer the harder question: does the language preserve the intended meaning, sound natural and fit the context in which a speaker would actually use it?

Thembi, a Dial8 venture in the data-review stage before its first fine-tuning experiment, is being built around that distinction. Its current product work turns language judgement into a consented, traceable workflow before model performance is allowed to become the headline.

Source quality is not fluent approval

The current Thembi repository starts with a licensed English-to-Xitsonga parallel corpus. The source record captures its owner, licence, permanent link and usage restrictions. Automated checks cover structure, length, alignment, duplication, placeholders and possible benchmark leakage.

Those checks matter. They show where an example came from and catch problems that software can identify consistently. They do not establish that a proposed translation is natural Xitsonga.

The product therefore separates three different claims:

• The source and permission have been checked.
• Automated quality checks have passed.
• A fluent speaker has approved the language.

The current candidate rows remain pending until the third claim is true. This prevents provenance, technical cleanliness and linguistic quality from collapsing into one reassuring but inaccurate status.

Make the judgement specific enough to use

“Please review this translation” sounds simple, but it leaves too much hidden. One reviewer may focus on spelling. Another may accept the meaning while disliking the register. A third may rewrite a sentence without recording what was wrong.

The Thembi review workspace presents one English prompt and one proposed Xitsonga version, then asks the reviewer to choose among four decisions:

• Approve the example.
• Provide a corrected version.
• Mark it understandable but unnatural.
• Mark the meaning as wrong.

Optional scores for meaning, naturalness and spelling preserve more detail. Notes create room for dialect, wording and context that a fixed rating cannot hold.

This is a product design decision, not only a data format. The interface must make a demanding linguistic task clear enough for a volunteer to complete while preserving evidence that can guide a later training decision.

Use small batches to protect attention

The current workflow assigns ten examples at a time. A reviewer can finish one batch, see their progress and decide whether to continue.

The number is not presented as a scientific optimum. It is a practical boundary in the working product. Language judgement requires sustained attention, and an endless queue would make the contribution feel abstract. A small batch gives the reviewer a visible unit of work and gives the project a complete set of decisions it can inspect.

Assignments also prevent the same person from reviewing the same item twice. Each item can target more than one review, which means coverage is not treated as a single opinion by default.

The product is designed to collect judgement without pretending that collecting more clicks automatically produces better language.

Disagreement is evidence, not a system error

Two fluent speakers can read the same sentence differently. They may prefer different wording, notice different context or disagree about whether an expression is natural.

Thembi does not need to hide that variation to make the dataset look clean. Its administration view is designed to surface items that have received different decisions after multiple reviews. Those items become an adjudication queue.

This creates a more honest path from contribution to training data:

1. A reviewer records an explicit decision. 2. The system preserves any correction, score and note. 3. A second judgement can confirm or challenge the first. 4. Disagreement remains visible for resolution. 5. Only reviewed evidence can move toward an approved dataset.

The disagreement is useful because it shows where context, register or the source sentence needs closer attention. Removing it too early would remove part of what the project needs to learn.

Keep evaluation outside the training loop

Thembi currently separates candidate training examples from a held-out benchmark. The benchmark covers translation in both directions, conversation, writing, summarisation, grammar, South African context and code-switching. Its private answers must not be included in training.

That boundary makes a later comparison meaningful. If the model has already trained on the material used to judge it, an improved score may only show familiarity with the test.

The repository also sets a narrow first model goal: compare the untouched base model with a small adapted version on approved Xitsonga tasks without damaging its English ability. The current plan is to review the data, establish the baseline, run a limited training experiment and compare the results before increasing the dataset.

This is evidence of a defined development method, not evidence that the adapted model already works. No trained Thembi model, performance improvement or public release is being claimed here.

Consent and provenance belong inside the product

The review portal requires a private access code and records how a contributor uses Xitsonga. It also asks for explicit consent to use corrections in training and evaluation, with public credit kept optional.

Those details cannot be postponed until after the language has been collected. They determine whether the project has permission to use a contribution and whether a reviewer may be named. The same principle applies to source material: public availability is not automatically permission to train a model.

For a community-led language product, trust depends on making those boundaries operational. The interface, database and data policy have to agree about what was contributed, who may use it and which status it has earned.

What progress means at this stage

Thembi has a working review portal, a structured review database, a documented source policy, prepared candidate data and a defined benchmark-and-training workflow. It is still before the first approved training experiment.

The next meaningful evidence is not a larger unreviewed dataset. It is completed fluent review, resolved disagreements, an untouched baseline and a controlled comparison using only approved training rows.

This is how Dial8 builds things that widen possibility: not by rushing an ambitious vision toward a polished claim, but by turning the uncertain part into a working product. For Thembi, the immediate possibility is a way for Xitsonga speakers to shape the evidence a future language model will learn from. The model comes after that foundation has earned trust.