Closed aittalam closed 3 months ago
LGTM once comments are addressed, 🔥
Thank you! The lm-harness entrypoint still fails tests but we can ignore it for now (it works when running integration tests locally, there might be a library version mismatch to fix)
Three main commits:
allow_remote_code
to config => allows us to run Phi modelsmax_samples
bug (now if the value is 0 all samples are considered, if it is larger than len(dataset) we get a warning and all the dataset is used)APIModelClient
class and made OAI and Mistral as two subclasses as so much code was in common)Tested manually with different configs (seq2seq models locally, OAI, Mistral) and with pytest.