resume script: fix dataset module names + opt-out of hf_transfer

run_resume_stage2.sh referenced data.build_holdout, but the actual
modules are train.make_sft_dataset and eval.make_holdout. Also
export HF_HUB_ENABLE_HF_TRANSFER=0 in both pipeline scripts so a
missing hf_transfer wheel can't take down model loading.

Made-with: Cursor

scripts/run_full_pipeline.sh

@@ -11,6 +11,11 @@
 #
 set -euo pipefail
 
+# HF Spaces flips on HF_HUB_ENABLE_HF_TRANSFER, but the matching wheel
+# is not always installed; opt out so AutoConfig/from_pretrained don't
+# crash when hf_transfer is missing.
+export HF_HUB_ENABLE_HF_TRANSFER=0
+
 echo "[1/6] Installing GPU stack ..."
 pip install -q --upgrade pip setuptools wheel
 # Python-3.10-safe `future`; legacy versions fail with

scripts/run_resume_stage2.sh

@@ -11,6 +11,11 @@
 # repo has been re-cloned. Safe to re-run.
 set -euo pipefail
 
+# Belt-and-suspenders: HF Spaces sets HF_HUB_ENABLE_HF_TRANSFER=1 which
+# requires the `hf_transfer` wheel; if any dep install drops it we'd
+# rather slow downloads than crash, so disable here.
+export HF_HUB_ENABLE_HF_TRANSFER=0
+
 # Reuse the install logic from run_full_pipeline.sh by sourcing only
 # steps 1 and 2. We'd rather duplicate a few lines than risk source-ing
 # a script that exits early if SFT adapter is missing.
@@ -28,7 +33,8 @@ pip install -q hf_transfer msgspec "torchao>=0.13.0" cut_cross_entropy || true
 pip install -q -r requirements.txt
 
 echo "[2/6] Building / verifying datasets ..."
-python -m data.build_holdout
+python -m train.make_sft_dataset --n 600 --out data/sft_train.jsonl
+python -m eval.make_holdout --out data/holdout.jsonl
 
 echo "[3/6] Skipped (SFT + Stage 1 already on Hub)."