Greg Tatum
Greg Tatum
I see the language mentioned in the NLLB paper, but I don't see it in the datasets: https://arxiv.org/pdf/2207.04672 https://opus.nlpl.eu/results/en&th/corpus-result-table
We need a good sanity check for when we deploy our new versions. We should have a integration test that covers the relationship between the add-on and our client. For...
This is a requirement for #2931. I can't quite get parity with a TypeScript implementation, which used more inference, as Flow has a bug in object type inference. I'm keeping...
### Description Building on macOS doesn't detect correctly that it is on an ARM processor. List of changes: - Adds a new if branch for ARM64. Added dependencies: none ###...
### Description I'm building on macOS arm64 using clang and encountered a bug with a compile error. Example error: ``` [ 50%] Building CXX object src/CMakeFiles/marian.dir/3rd_party/ExceptionWithCallStack.cpp.o /Users/greg/dev/marian-dev/src/3rd_party/ExceptionWithCallStack.cpp:205:30: error: variable length...
### Description On an Apple arm64 build, I got an error running cmake. `CPUINFO` is a blank string. Since it's not in quotes, the argument list is too small creating...
A large class of errors in the pipeline on the H1-2025 run is that of language tag matching. Our config only supports ISO 639-1 language tags with the `src` and...
From Matrix: > Greg: Looks like en-hu has problems with short words. I've been investigating and it looks like it's learned to add things like (angol nyelven) to the end...
In `llm.py` I batch multiple translations into a single request. `input = translations_batch_to_text(translations)` > evgeny: Did you test batching vs using a single example per call? I suspect batching is...
In #1209 I'm adding LLM evaluation of our translations with explanations. We can also make it work for translations without a reference. > @evgenyrp stated: We can also do referenceless...