update AutoGuess template

[tsite]

fix rwkv bug
add vicuna, dots llm

[LostRuins]

did you test the vicuna one? I have not found a single model that matches it. if its not a real template in use we should not add it.

[tsite]

wizardlm-2 (https://huggingface.co/alpindale/WizardLM-2-8x22B) uses a variant of Vicuna - I updated the AutoGuess template to match

[LostRuins]

Also what's wrong with the rwkv template, that was from @henk717

Generally for these weird and uncommon formats I feel its not ideal to clutter the autoguess with something that literally nobody uses except for 1 tiny case, vicuna is basically unused in the wild and when necessary alpaca outperforms it in all cases.

[henk717]

Rwkv should remain as is. It was directly given to me by one of the rwkv developers so only they should have the say on it. Just like with Mistral i like to keep the official templates official.

[tsite]

The rwkv template in autoguess.json does not match the one in https://github.com/BlinkDL/ChatRWKV/blob/main/API_DEMO_CHAT.py. Will close this pr out as the added templates are for "uncommon" models, but I recommend fixing the rwkv template in a separate pr.

[kallewoof]

+1 on fixing the RWKV chat templates. This is what the ChatRWKV demo code looks like:

        out = run_rnn("User: " + msg + "\n\nAssistant:")

where run_rnn simply tokenizes the first arg and passes it to the model, so at least in this case, the format is "User: " for the user prefix and "Assistant:" (no space after ":") for the assistant. If @henk717 could verify that would be a plus.