| # Gemma 4 31B data: RULER (re-fetch generator), RULER 6 tasks x 4K/32K/128K with the Gemma tokenizer, MRCR bins, DocVQA link. | |
| set -x | |
| source ~/kvq-env/bin/activate | |
| cd ~/datasets | |
| [ -d ruler/RULER ] || git clone --depth 1 https://github.com/NVIDIA/RULER.git ruler/RULER | |
| (cd ruler/RULER && git log -1 --format="RULER commit %h %cd") | |
| cd ruler/RULER/scripts/data/synthetic/json | |
| [ -f PaulGrahamEssays.json ] || python download_paulgraham_essay.py | |
| [ -f squad.json ] && [ -f hotpotqa.json ] || bash download_qa_dataset.sh | |
| ls -la | |
| cd ~/bench | |
| export KVQ_TOK=~/models/gemma-4-31B-it KVQ_DATA=data_gemma | |
| mkdir -p data_gemma | |
| ln -sf ~/bench/data/docvqa_1000.jsonl data_gemma/docvqa_1000.jsonl | |
| KVQ_RULER_TASKS=niah_single_2,niah_multikey_1,niah_multikey_3,niah_multivalue,vt,qa_2 KVQ_RULER_LENGTHS=4k,32k,128k \ | |
| python prep_data.py --stage ruler --workers 18 | |
| python make_6t.py | |
| python mrcr.py prep | |
| echo PREP_GEMMA_DONE | |