| # FAST 7B verifier training (~30-45 min) via sequence packing + FlashAttention-2. | |
| # Packing concatenates the short rows (p50=602) into dense 2560-token sequences, | |
| # cutting ~8127 padded rows to ~2000 dense ones -> far fewer steps. Packing REQUIRES | |
| # FlashAttention-2 (else it attends across packed examples and corrupts training): | |
| # pip install flash-attn --no-build-isolation # ~5-15 min build on Colab | |
| # python -m mathcompose.train.train --task v --config configs/verifier_v_7b_fast.yaml \ | |
| # --gpu-preset none --push --hub-model-id <you>/mathcompose-verifier-7b | |
| extends: "verifier_v_7b.yaml" | |
| model: | |
| attn_implementation: "flash_attention_2" | |
| train: | |
| packing: true | |