Buckets:
| # Procedural testbed arm: run_proc_arm.sh <name> <train flags...>; trains on 32k procedural items, evals proc test + dev. | |
| cd /workspace/think-decide | |
| export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True | |
| NAME=$1; shift | |
| NT=0; KMAX=1 | |
| for f in "$@"; do case $prev in --n-think) NT=$f;; esac; prev=$f; done | |
| [ "$NT" != "0" ] && KMAX=8 | |
| P=/workspace/data/procedural/canonical | |
| python src/train.py --out runs/$NAME --train-file $P/train.jsonl --dev-file $P/dev.jsonl --train-limit 32000 \ | |
| --max-length 8192 --lr-bb 2.5e-5 "$@" > logs/${NAME}_train.log 2>&1 | |
| for s in proc_test proc_dev; do | |
| python src/eval.py --model runs/$NAME/model.safetensors --n-think $NT --kmax $KMAX --split $s --compile \ | |
| --out runs/$NAME/$s >> logs/${NAME}_eval.log 2>&1 | |
| done | |
| grep -E "eval_done|proc all" logs/${NAME}_eval.log | |
Xet Storage Details
- Size:
- 813 Bytes
- Xet hash:
- ec859bf097ce8a0971651c2b9bd301dd3bc8b4aa67457086d6d1b325f5bee9cd
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.