The Dataset Viewer has been disabled on this dataset.

V2-LQ pretraining pool (tokenized, Megatron format)

This is the exact training data of v2-lq-seed1234/ in anonymous-md/test. It holds 80,610,628 documents from anonymous-md/ProseOnlyRepair_linguistic_LQ, tokenized with the tokenizer that ships with those checkpoints.

File Size (bytes) SHA-256
post_text_document.bin 138,127,611,356 a33e605759215a20820e8e81b1e35eacf76cf937363e41e38241b24ba0dc7160
post_text_document.idx 1,612,212,602 1fe149bb7873632da098384c7a800b58cbbe82254b8f1a10aea4e35c8b16cb6c
selection.json 150,135

Format

  • Megatron-Core indexed dataset (MMIDIDX), token IDs stored as int32.
  • 80,610,628 documents and 34,531,902,839 tokens. Every document ends with <|endoftext|> (ID 151643), so 34,451,292,211 of the tokens are text.

How the documents were selected

  • The source has 2,619 parquet files. They were shuffled with seed 42 and read whole, in that order, until the pool reached the 34.4B-token target. That happened after 2,617 files; selection.json lists the files used and their token counts.
  • Rows with unrealized or had_error set, or with an empty text, were skipped (22,107,388 rows).

Training settings that used this pool

v2-lq-seed1234/ was trained with Megatron-Bridge 0.4.0rc0 and Megatron-Core 0.17.0rc0, for 16,406 steps of 512 sequences, with these dataset settings:

dataset:
  blend:
  - - /path/to/post_text_document   # this folder's files, without the extension
  - - 1.0
  split: 999,1,0
  sequence_length: 4096
  random_seed: 1234

Megatron-Core builds its sample order deterministically from the pool, these settings and the step and batch counts. A run with the same versions and settings that changes only the weight-initialization seed therefore sees the same batches, in the same order, as v2-lq-seed1234/.

Download

hf download anonymous-md/v2-lq-pool --repo-type dataset --local-dir ./v2-lq-pool
Downloads last month
21