how many commoncrawl snatpshot this work used?
hi, thank you in advance! I have two questions.
how many snapshots does this work use ? so far, the number of available commoncrawl snapshot is approximately 90. Did you guys use all snapshots and do steps(url filtering, text extraction , language identification, repetition removal and more deuplication process; reference link[1] here: https://the-decoder.com/falconlm-open-source-language-model-beats-metas-llama/ )
which data type does this work use(e.g, WET format, WARC foramt) it seems you guys use "trafilatura " as stated in data description. For me, it seems to use WARC format. However, above reference link[1]'s figure tells that text extraction rarely discards the text. However, if using WARC format, a comparable amount of text should be discarded. So could you specify which format you guys choose?
Thank you!
Hey!
- To keep the data homogeneous and include both older & recent content we sample from all CommonCrawl dumps available, but only keep a few segments from each. The
dumpandsegmentfields on each instance contain the relevant information. - We indeed start from
.WARCfiles as we found the text extraction used in.WETfiles to be of relatively poor quality. The figure is a bit misleading (we will add some clarification in the paper), but at this stage it only measures removal rate in number of documents -- so this corresponds to ~2% of docs for which we failed to extract content. Obviously if this was expressed in tokens/words, the removal rates would be much higher at this stage.
@FalconLLM Could you provide more details about the perceived lower quality of WET files? Are you suggesting that the extraction of content from .wet files is flawed in your assessment?