spb/wpai0002_uqo

1#!/usr/bin/env bash2# Instance 3 GPU $1: STRAP-style inverse-paraphrase model on a pretrained English seq2seq backbone (google/flan-t5-large, 780 M;3# facebook/bart-large gives NaN gradients from the 2nd step in bf16 AND fp32 with torch 2.11 — bart-base trains fine — so T5 v1.1),4# fine-tuned on the same Medium pairs/splits, then val/test/held-out predictions. Exploratory comparison system "B2_bart-large".5set -uo pipefail6cd ~/wpai00027export PATH="$HOME/.local/bin:$PATH" CUDA_VISIBLE_DEVICES=${1:-0} PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True HF_HUB_DISABLE_PROGRESS_BARS=18P=data/pairs/medium_v1_haiku/validated/pairs.jsonl; S=data/splits/medium_v1/splits.jsonl9.venv/bin/python scripts/train_b2_hf.py --pairs $P --splits $S --run medium_b2_flan_t5_large --model-name google/flan-t5-large --max-steps 20000 --batch-size 12 --max-len 448 --lr 1e-4 --warmup 500 --eval-every 1000 --patience 6 --device cuda --system B2_flan-t5-large --predict-split val --pred-limit 3000 --pred-splits test,heldout_domain,heldout_period --pred-out artifacts/predictions > train_t5large.log 2>&110echo "CHAIN t5large DONE" >> chain_done.txt11