When the training model is successfully restarted, the start-up service fails to find the model. about cakechat

[nicolas-ivanov]

@Spring1212, the size of vocabulary that you are trying to use is only 26 tokens - not good. Apparently you rewrote the default vocabulary file by running prepare_index_files.py. See the corresponding section of our Readme for a fix guideline: https://github.com/lukalabs/cakechat#training-the-model-from-scratch

[Spring12111]

Is this due to too little training data?

[Spring12111]

Can adding training data avoid this problem?

[nicolas-ivanov]

Is this due to too little training data?

You got a different index files because apparently you run prepare_index_files.py and it rewrote the original index files. Even if you had training data in abundance, running this script would still result in overwriting index files, so won't be able to use our pretrained model.

If you just want to use our pretrained model without finetuning on you own data, you don't need to run prepare_index_files.py script. Restore the default index files following the link I send you in the previous message and run the server with our pretrained model after that.

And in case you do want to tune our model or to train your own from scratch, please, read carefully the corresponding sections from our Readme: https://github.com/lukalabs/cakechat#training-the-model

[Spring12111]

According to the self-report file training model from scratch, every step of adding training data into the file is operated.

[Spring12111]

Is this still a pre-training model?

[nicolas-ivanov]

Seems like I got the problem: the server is trying to use a reverse model for sampling_reranking mode of generating answers and, I assume, you didn't train it, since it's not mentioned in this section of our Readme.

Option 1: change sampling_reranking to sampling here: https://github.com/lukalabs/cakechat/blob/master/cakechat/api/config.py#L4

Option 2: run python tools/train.py -r to train the reverse model.

In either case, please, note that you'll need to use a larger training set with larger vocabulary to get any meaningful results after training.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.