Rishi Jonathan Rebello

From the [marian tokenizer](https://github.com/huggingface/transformers/blob/e342ac7e0390d157e8c14a8a7c5cf2e715341cee/src/transformers/models/marian/tokenization_marian.py#L146), the bos_token_id is not initialized. Instead it recommends using the decoder_start_token_id from the config. For this model, the decoder_start_token_id is [65000](https://huggingface.co/Helsinki-NLP/opus-mt-en-zh/blob/main/config.json#L24). Example: ``` session = InferenceSession("opus-mt-en-zh-onnx-301/model.onnx")...

Alternatively, I found that the optimum library makes working with seq2seq models in ONNX much easier. ``` from transformers import AutoTokenizer, pipeline from optimum.onnxruntime import ORTModelForSeq2SeqLM model_path = "Helsinki-NLP/opus-mt-en-zh" tokenizer...

@nvoxland I tried the clear_cache() method and it did not work. It still only shows the stale data and does not see the new commit to the dataset.