EdoardoPona/predicting-inductive-biases-RL

Load custom transformer in RL4LM

Closed this issue · 2 comments

EdoardoPona commented

RL4LM loads the models from the config file by using huggingface's AutoModel, we need to do one of the following

  • code a transformer with hf (compatible with RL4LM). This has to be a sequence model, so a Decoder.

  • save an untrained model locally that can be loaded by RL4LM or

  • upload this model on the hub, so that it can be loaded directly from there

EdoardoPona commented

A similar process has to be done for the tokenizer

EdoardoPona commented

@diogo-cruz is the work on the tokenizer finished with commit feff6a7?
If so, we can close this.