Explain the language goal without claiming the result
Pino SI’s public site describes adapting an open model for Tagalog, English and Taglish, then extending to more Philippine languages over time. It says native-speaker-written Taglish questions will help test the result. Evaluation is still marked as planned.[1]
This is Kooya’s account of its own planned research, not a report of achieved scores. Naming a language in a mission does not establish reliable capability in it.
Test mixed language on its own terms
Zhang and colleagues’ 2023 study examines code-switching across several tasks and finds that multilingual ability does not automatically imply proficiency on mixed-language text. Those findings concern the systems and conditions studied in that paper. They do not measure Pino SI or establish its future performance.[2]
Our research interpretation is straightforward: an English test and a Tagalog test cannot fully stand in for the way a Taglish request is expressed. A speaker can move between languages while keeping one thought, and the important constraint may appear on either side of that switch.
A fictional coding request illustrates the issue: “Sort by date, pero huwag baguhin ang original list.” The crucial requirement is to preserve the input. An answer could use friendly Taglish yet still mutate the list. That would be a task failure despite a superficially natural response.
Ask native speakers to preserve meaning
Native-speaker participation can help establish whether a question sounds like something a person would actually ask and whether an answer carries the intended meaning. In our evaluation rationale, that contribution goes beyond checking grammar or correcting spelling.
Consider a question with an unstated subject, an informal abbreviation or a polite phrase around a firm instruction. The evaluator needs to decide what a reasonable speaker would understand, and whether ambiguity should lead to clarification. Mechanical translation alone may leave that judgment unresolved.
Separate naturalness from correctness
The public plan asks whether Taglish improves while English remains strong, and uses coding as a first pilot because generated answers can be run. These are evaluation intentions rather than published findings.[1]
Our interpretation is that several judgments belong side by side. Did the model understand the instruction? Is its explanation natural for the intended reader? Does the answer actually solve the task? A single overall impression can hide an important failure in one of these dimensions.
For the fictional sorting request, a reviewer could first inspect the preservation requirement, then execute the code, then assess the explanation. If the solution works but the wording is awkward, that is a different improvement problem from a fluent answer that changes the wrong data.
Treat regional coverage as a sequence of evidence
The regional-languages and knowledge focus is described as a foundation of Pino SI’s research roadmap. The site places broader language expansion over time, rather than presenting every Philippine language as already supported.[1]
Our position is that evidence should be specific enough for a reader to understand its limits. A result in one language, one task or one style of writing should not silently become a claim about all regional languages. Nor should a few successful greetings stand in for sustained understanding.
Before a broad capability statement would be justified, the relevant test would need to represent the language and use case in question. The point is not to announce a new testing program here; it is to explain why the project’s current wording remains deliberately prospective.
Keep local knowledge distinct from linguistic fluency
A response can use familiar vocabulary while relying on an incorrect or outdated fact. Conversely, a technically correct statement may need clearer language to be useful. In Kooya’s research rationale, these are related but separate evaluation questions.
For future advisory and public-service tools, the public plan describes answering from approved sources, citing them and using human review where appropriate. That is a planned approach to current information, not a claim that a public advisory system has already been deployed.[1]
The practical implication is to inspect both the answer and its support. The model should not receive credit for cultural familiarity when the factual claim is unsupported. Where a question is underspecified, asking for context may be a better outcome than producing a polished guess.
Our public-service pilot note shows how source freshness and local context would matter in a farming advisory or delivery question.
Describe progress in terms a speaker can recognize
For a person who might eventually use the technology, “better language performance” should connect to observable behavior: fewer misunderstood constraints, clearer explanations or a more appropriate clarification question. Those are examples of useful outcomes, not results we are reporting today.
Pino SI’s regional ambition will be meaningful to the extent that evidence eventually supports the language and task claims made about it. At this stage, the honest description remains a research effort seeking that evidence, authored by the organization undertaking the work.