Upload scripts/dl_eval32k.sh with huggingface_hub
Browse files- scripts/dl_eval32k.sh +37 -0
scripts/dl_eval32k.sh
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
# Pull exactly what the 32k eval needs. Explicit --include on every call:
|
| 3 |
+
# an unfiltered download of this repo once pulled 167 GB and broke a build.
|
| 4 |
+
set -uo pipefail
|
| 5 |
+
cd "$(dirname "$0")"
|
| 6 |
+
export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
|
| 7 |
+
say(){ echo "[$(date -u +%H:%M:%S)] $*"; }
|
| 8 |
+
mkdir -p checkpoints_680m/rank_0 checkpoints_bluscriptcp/rank_0 Data_Loader/Data/b2_8192_16384 Data_Loader/Data/evalsrc
|
| 9 |
+
|
| 10 |
+
say "1/4 varybatch.ckpt (the untrained base -- the control)"
|
| 11 |
+
hf download navingv/varybatch --quiet --include 'varybatch.ckpt' --local-dir dl_vb || exit 1
|
| 12 |
+
mv -f dl_vb/varybatch.ckpt checkpoints_680m/rank_0/ && rm -rf dl_vb
|
| 13 |
+
|
| 14 |
+
say "2/4 p1final step1200 rank0 (5.14 GB)"
|
| 15 |
+
hf download unparallelled/BluScriptCP --quiet --include 'checkpoints/p1_680m_step1200_rank0.ckpt' --local-dir dl_p1 || exit 1
|
| 16 |
+
mv -f dl_p1/checkpoints/p1_680m_step1200_rank0.ckpt \
|
| 17 |
+
checkpoints_bluscriptcp/rank_0/blumodelcp_run12_step_1200.ckpt && rm -rf dl_p1
|
| 18 |
+
|
| 19 |
+
say "3/4 haystack -- MUST be the same file phase 1 used or the numbers are not comparable"
|
| 20 |
+
hf download unparallelled/BluScriptCP-data --repo-type dataset --quiet \
|
| 21 |
+
--include 'cp_data_by_length/b2_8192_16384/val_shard_00000.bin' --local-dir dl_hay || exit 1
|
| 22 |
+
mv -f dl_hay/cp_data_by_length/b2_8192_16384/val_shard_00000.bin \
|
| 23 |
+
Data_Loader/Data/b2_8192_16384/ && rm -rf dl_hay
|
| 24 |
+
|
| 25 |
+
say "4/4 a train split for the loader -- the trainer builds one even for pure generation"
|
| 26 |
+
hf download unparallelled/BluScriptCP-data --repo-type dataset --quiet \
|
| 27 |
+
--include 'cp_data_by_length/b1_4096_8192/train_shard_00000.bin' --local-dir dl_tr || exit 1
|
| 28 |
+
mv -f dl_tr/cp_data_by_length/b1_4096_8192/train_shard_00000.bin Data_Loader/Data/evalsrc/ && rm -rf dl_tr
|
| 29 |
+
cp -f Data_Loader/Data/b2_8192_16384/val_shard_00000.bin Data_Loader/Data/evalsrc/
|
| 30 |
+
|
| 31 |
+
say "=== sizes ==="
|
| 32 |
+
stat -c ' %s %n' checkpoints_680m/rank_0/varybatch.ckpt \
|
| 33 |
+
checkpoints_bluscriptcp/rank_0/blumodelcp_run12_step_1200.ckpt \
|
| 34 |
+
Data_Loader/Data/b2_8192_16384/val_shard_00000.bin \
|
| 35 |
+
Data_Loader/Data/evalsrc/train_shard_00000.bin 2>/dev/null
|
| 36 |
+
say "expected: varybatch 1362674564 | p1final 5141244514 | haystack 199754694"
|
| 37 |
+
say "DOWNLOAD DONE"
|