Datasets:
What is the total # tokens after sampling proportion? 1.7T or 1.65T
Hi! Thanks for sharing the dataset and sampling proportion!
I noticed one discrepancy on token counts. On the data card, it says:
A subset of total data was used for training of OLMo 7B-v1.7. The token counts are based on the full dataset, whereas taking into account sampling proportion gives the final actual token counts used for training --- 1.715 trillion tokens.
However, when calculating the sum of tokens based on the sampling ratio, the total # tokens is 1.65T. There is a gap of 70B tokens, which is about the size of C4 with sampling proportion.
Am I miss anything? Is the calculation correct, or the sampling proportion needs to be updated?
Thanks!
Hello, Sorry for late reply! Briefly, we sampled C4 at 100%, not 50%.
Exact counts as shown below
| source | billion tokens | type | upsample | final |
|---|---|---|---|---|
| dolma: gutenberg books | 5.3 | REF | 100% | 5.3 |
| dolma: pes2o | 57.2 | REF | 100% | 57.2 |
| dolma: wikipedia & wikibooks | 3.7 | REF | 200% | 7.4 |
| redpajama: stackexchange | 19.6 | REF | 100% | 19.6 |
| redjapama: arxiv | 28.0 | REF | 100% | 28.0 |
| proofpile2: algebraic stack | 12.6 | REF | 100% | 12.6 |
| proofpile2: openwebmath | 12.7 | REF | 100% | 12.7 |
| tulu: flan v1 (v1-decontaminated-60M-shots_all-upweight_1-dialog_true-sep_newline) | 16.5 | REF | 100% | 16.5 |
| CC News | 14.3 | REF | 100% | 14.3 |
| dolma: c4 | 138.4 | HQW | 100% | 138.4 |
| dolma: reddit | 79.9 | HQW | 100% | 79.9 |
| refinedweb | 456.4 | HQW | 100% | 456.4 |
| megawika v1 (refs from wikipedia) | 4.6 | REF | 100% | 4.6 |
| starcoder | 263.8 | C | 100% | 263.8 |
| dolma: cc high | 356.8 | W | 50.2% | 179.2 |
| dolma: cc middle | 452.4 | W | 50.4% | 227.8 |
| dolma: cc low | 386.3 | W | 49.6% | 191.4 |
| total | 1715.1 |
Thanks for clarifying, @soldni ! That makes sense.
I got 50% sampling proportion from this page: https://huggingface.co/datasets/allenai/dolma#summary-statistics-v17
Shall it be corrected?