Datasets:
depth int64 | aspect_ratio int64 | head_dim int64 | mlp_multiple int64 | vocab_size int64 | seq_len int64 | num_gpus int64 | batch_size int64 | token_multiple int64 | rope_theta int64 | weight_decay float64 | dtype string | lr float64 | data_mix dict | tokenizer string | filename string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
30 | 64 | 128 | 4 | 32,768 | 2,048 | 8 | 64 | 20 | 10,000 | 0.05 | bfloat16 | 0.001 | {
"fineweb-edu-100BT.shuffle.bin": 0.96875,
"smoltalk.shuffle.bin": 0.03125
} | tokenizer_50M.bpe | llm.d30.2.pt |
LLM Speedrun token streams
Pre-tokenized training artifacts for the LLM speedrun exercises.
| File | Description | Tokens |
|---|---|---|
tokenizer_50M.bpe |
JSON-serialized BPE tokenizer | — |
fineweb-edu-10BT.shuffle.bin |
Shuffled FineWeb-Edu sample/10BT token stream |
9,440,023,113 |
smoltalk.shuffle.bin |
Shuffled SmolTalk data/all token stream |
875,269,408 |
The .bin files are headerless, little-endian unsigned 16-bit token IDs and can be memory-mapped with NumPy:
from huggingface_hub import hf_hub_download
import numpy as np
path = hf_hub_download(
"zkolter/llm_speedrun",
"fineweb-edu-10BT.shuffle.bin",
repo_type="dataset",
)
tokens = np.memmap(path, dtype="<u2", mode="r")
Use tokenizer_50M.bpe with the matching LLM speedrun BPE implementation. See DATA_CARD.md for provenance, processing, licensing, and limitations.
- Downloads last month
- 154