fam/nanochat-omni
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Big DataLoader refactor: BOS-aligned dataloaders with epoch tracking for pre/mid-training

Browse Source 
The new DataLoader ensures that every token sequence in train/val batches has a BOS token
at the beginning. Therefore, no token streams start abruptly in the middle of a document,
which could be confusing for the model. Note that this changes the loss scale because there
are fewer confusing tokens in the train/val batches. The main downside is that we now waste
about 35% of tokens due to cropping. This is ok because we have a lot of data. See dev/LOG.md
entry for this change for a lot more information.
This commit is contained in:
Andrej Karpathy
2026-01-13 20:05:47 +00:00
parent 23985413aa
commit 43c29dd9d5
7 changed files with 330 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files
speedrun.sh
+5 -3
View File
@@ -55,8 +55,8 @@ python -m nanochat.report reset
# each shard is ~100MB of text (compressed), so this is about ~800MB of data on disk
python -m nanochat.dataset -n 8
# Immediately also kick off downloading more shards in the background while tokenizer trains
# See comment below for why 240 is the right number here
python -m nanochat.dataset -n 240 &
# See comment below for why 370 is the right number here
python -m nanochat.dataset -n 370 &
DATASET_DOWNLOAD_PID=$!
# train the tokenizer with vocab size 2**16 = 65536 on ~2B characters of data
python -m scripts.tok_train --max_chars=2000000000 --vocab_size=65536
@@ -70,7 +70,9 @@ python -m scripts.tok_eval
# Chinchilla says #tokens = 20X #params, so we need 561e6 * 20 = 11.2B tokens.
# Assume our tokenizer is 4.8 chars/token, this is 11.2B * 4.8 ~= 54B chars.
# At 250M chars/shard, this is 54B / 250M ~= 216 shards needed for pretraining.
# Round up to 240 for safety. At ~100MB/shard, this downloads ~24GB of data to disk.
# Round up to 240 for safety. Also, the new DataLoader wastes about 35% of tokens to cropping
# so 240 / (1 - 0.35) = 370 shards are needed.
# At ~100MB/shard, this downloads ~37GB of data to disk.
# (The total number of shards available in the entire dataset is 1822.)
echo "Waiting for dataset download to complete..."
wait $DATASET_DOWNLOAD_PID