Scriptwriter dataset sync

restart-train.sh

@@ -25,6 +25,7 @@ export HF_DATASET_REPO="${HF_DATASET_REPO:-datamatters24/scriptwriter-corpus-ia}
 export HF_MODEL_REPO="${HF_MODEL_REPO:-datamatters24/scriptwriter-lora-ia}"
 export BASE_MODEL="${BASE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}"
 export RUNPOD_SKIP_FORCE_PIP=1
+export CUDA_VISIBLE_DEVICES=0
 export MAX_SEQ_LENGTH=2048
 export MAX_SEQ_LENGTH_CAP=2048
 export PER_DEVICE_TRAIN_BATCH_SIZE=1
@@ -65,7 +66,7 @@ export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
 export RESUME_FROM_CHECKPOINT=1
 exec > >(tee -a /workspace/logs/train_live.log) 2>&1
 echo "========== TRAIN START $(date -Is) =========="
-echo "OOM-safe: seq=2048 batch=1 accum=8 gpu0-only + checkpointing"
+echo "OOM-safe: seq=2048 batch=1 accum=8 CUDA_VISIBLE_DEVICES=0"
 ls /workspace/models/lora/checkpoint-* 2>/dev/null | tail -5 || echo "no prior checkpoint"
 nvidia-smi || true
 python3 scripts/train_runpod.py

scriptwriter-runpod-ready.tgz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:acd25ad6f3b274dda2d69d44438dea32bf84d60440d418453194cf34c2402f70
-size 10148
+oid sha256:e5d2813d7c694fd3e3a48474e5660c76faf114720a5be44eecd4e92afe723e28
+size 10140