summaryrefslogtreecommitdiff
path: root/ep_run/prepare_fineweb.py
diff options
context:
space:
mode:
authorOscar Wan <oscarwan@oriong12.stanford.edu>2026-07-21 15:28:06 -0700
committerOscar Wan <oscarwan@oriong12.stanford.edu>2026-07-21 15:28:06 -0700
commitad4622e42abf2a55a37c5e522a6f8302e2e95cfd (patch)
treece668b916f9f16c4153c8b658bd772cd338ff90b /ep_run/prepare_fineweb.py
parent09c52fa97a946cf6dbdc500aef88da8522c2ce7d (diff)
Add portable 135M BP sweep handoff
Match the active EP configuration and support four-GPU BP candidates from collaborators' data locations. Co-authored-by: Cursor <cursoragent@cursor.com>
Diffstat (limited to 'ep_run/prepare_fineweb.py')
-rw-r--r--ep_run/prepare_fineweb.py5
1 files changed, 3 insertions, 2 deletions
diff --git a/ep_run/prepare_fineweb.py b/ep_run/prepare_fineweb.py
index 33fce59..afd2e45 100644
--- a/ep_run/prepare_fineweb.py
+++ b/ep_run/prepare_fineweb.py
@@ -11,7 +11,7 @@ Phases (all resumable-ish, markers for the watcher):
Docs are joined with a <|eot|> separator (id 0). vocab 32768 fits uint16.
NFS note: peak disk = raw parquet ~28GB + bins ~20GB; keep raw/ for tokenizer reruns.
"""
-import pickle, time
+import os, pickle, time
from pathlib import Path
import numpy as np
import pyarrow.parquet as pq
@@ -22,7 +22,8 @@ from tokenizers.trainers import BpeTrainer
from tokenizers.pre_tokenizers import ByteLevel
from tokenizers.decoders import ByteLevel as ByteLevelDec
-D = Path('/home/yurenh2/ept/ep_run/data/fineweb_edu')
+DATA_ROOT = Path(os.environ.get('EPT_DATA_ROOT', Path(__file__).resolve().parent / 'data'))
+D = DATA_ROOT / 'fineweb_edu'
RAW = D / 'raw'
D.mkdir(parents=True, exist_ok=True)
VOCAB = 32768