The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
dclm-100b as one GPT-NeoX token stream
HuggingFaceFW/dclm_100BT-shuffled (revision 2fa015e4044e, published order, no
shuffle) tokenized with EleutherAI/gpt-neox-20b (revision
c292233c833e, 50,279 tokens;
50277/50278 are the ChatML markers,
0 ends every document) into one contiguous uint32 stream:
116,670,862,301 training tokens in 435 shards of 2^28
tokens (the last shard may be shorter), after a held-out slice of
99,998,042 tokens in 76,129 documents. The holdout
requests 100,000,000 tokens from the stream's first documents and
ends before the document that would cross that limit. Training begins with
that next whole document; no document is split between validation and training.
The total stored stream has 116,770,860,343 tokens
and occupies 436 GiB. All 89,269,902 source document positions were processed. The store contains 89,269,902 nonempty documents; empty source texts emit no tokens.
The 100,000,000-token holdout target changes the training offset relative to stores built with the former 30M-token target. Training shards from those stores are not byte-identical prefixes of this split.
The upstream data is ODC-By 1.0 (attribution: the DCLM authors and Hugging Face's shuffled 100BT sample); this derivative carries the same license.
Layout (the delta-feedback-experiment document-stream-v1 store):
| File | Contents |
|---|---|
meta.json |
source, tokenizer, build packages (transformers 5.17.0, tokenizers 0.23.2, huggingface-hub 1.31.0, pyarrow 25.0.1), counts |
source.json |
the document universe: every parquet file's row groups |
val.bin, val.docs.npy |
the held-out slice and one (start, source) record per document |
train.NNNN.bin |
uint32 shards, one contiguous stream, 2^28 tokens each |
train.docs.npy |
one (start, source) record per training document |
dclm-100b.sha256 |
checksums of every file above |
Row r of the training stream is tokens [r·(L+1), (r+1)·(L+1)) for a model
of sequence length L, so any prefix of the shards is a training set by
itself: hf download a9lim/dclm-100b-neox --type dataset --include 'train.00[0-1]*.bin' --include 'val.*' --include '*.json' --include '*.npy' gives the first 20.
Shorter builds using the same source and tokenizer pins, build packages,
ordering, seed, and held-out target have byte-identical training prefixes.
- Downloads last month
- 69