Some issue about ICL Experience about autocompressors HOT 3 CLOSED

Hi @broalantaps!
No, we did not use any ICL data while fine-tuning any of the AutoCompressor models.

Regarding your second question, I suspect that you are referring to Table 4's caption in our paper.

...note that for BoolQ and MultiRC, demonstrations are too long to fit into 150 tokens.

For datasets like BoolQ and MultiRC were even individual demonstrations exceeded 150 tokens in length, we simply conducted a zero-shot evaluation where the AutoCompressor was prompted to answer the test example expressed as plaintext.

I hope this answers your questions!

from autocompressors.

Understand it, thanks for your replying~

from autocompressors.

Another question about Llama7b's version. What did you use between meta-llama/Llama-2-7b-hf and meta-llama/Llama-2-7b-chat-hf? Thansk~

from autocompressors.