Supported models & Tasks
Last updated
Last updated
These docs are outdated! Please check out for the latest information on the TitanML platform. If there's anything that's not covered there, please contact us on our .
- sequence_classification
or glue.
Note that , the General Language Understanding Evaluation benchmark, is a collection of sequence classification tasks used to evaluate natural language understanding systems - so you'd have to specify which columns in your dataset contain the sequences which are to be classified, as well as how many labels there are. Using glue
as the task and the glue task as the dataset is a handy shortcut.
- question_answering
The most common datasets for question answering are the , but TitanML does support others. If you decide to use a different training dataset, you must indicate to Iris whether your dataset contains unanswerable questions (see how to do this here).
- token_classification
TitanML also supports classification tasks involving individual tokens (including Named Entity Recognition). As with sequence classification, you must indicate which columns in your input dataset are to be classified, and how many labelled classes they are to be classified into.
- language_modelling
Causal language modelling is how large language models like GPT-4 and Claude are trained. The model learns to predict the next word (technically, token) given a string of previous words (tokens). TitanML supports language modelling for LLMs like OPT and pythia. Large models will automatically use state of the art parameter efficient training. See below for supported models.
(sequence to sequence) - language_modelling
TitanML also supports conditional language modelling as a task. Conditional language modelling (also known as sequence to sequence modelling) involves producing output tokens conditioned on both previous tokens, and an additional sequence. Examples include translation, & summarization. Provide language_modelling
as the iris task, and the platform will automatically deduce the task type from the model used. See below for supported models.
Supported models for language modelling
t5 family (conditional language modelling) (google/t5-efficient-small-el16
, t5-base
, google/flan-t5-base
, ...)
pythia family (language modelling) (EleutherAI/pythia-160m
, EleutherAI/pythia-410m
, ...)
OPT family (language modelling) (facebook/opt-125m
, facebook/opt-1.3b
, facebook/opt-350m
, ...)
Anything supported by huggingface's AutoModelForSequenceClassification
can be finetuned on the TitanML platform.
Anything supported by huggingface's AutoModelForTokenClassification
can be finetuned on the TitanML platform.
Anything supported by huggingface's AutoModelForQuestionAnswering
can be finetuned on the TitanML platform.