can we replace https://the-eye.eu/public/AI/pile/val.jsonl.zst

[luchangli03]

I can't download the file "https://the-eye.eu/public/AI/pile/val.jsonl.zst" in get_calib_dataset, can we use other data to replace it?
Thanks a lot.

[tonylins]

Hi, you can swap in other general-domain corpus and see how it works. AWQ should be quite robust to the distribution, but the final results might be a bit different.

[abhinavkulkarni]

In fact, the entire Pile dataset is not available on the website. I'm getting 404 error. Not sure temporary or permanent.

Not sure about any mirrors either.

[abhinavkulkarni]

Pile dataset has been uploaded again, so the AWQ should work.

In case it disappears again, you can replace

dataset = load_dataset("json", data_files="https://the-eye.eu/public/AI/pile/val.jsonl.zst", split="train")

with

dataset = load_dataset("imdb", "plain_text", split="train")

The idea is to pipe a small portion of training data through the LLM to search for optimal AWQ parameters. I'm sure the IMDB reviews are used as training data by most of these big LLMs.