Buckets:
| # At checkpoint-40: stop the 1e-5 run by session name, continue 20 steps at 2.5e-5 from its adapter (fresh optimizer, | |
| # unseen prompts only), then run the post-RL eval on the final adapter. | |
| cd /workspace/decision-reasoner/runs/v2 | |
| C=checkpoints/rl_grpo/checkpoint-40 | |
| until [ -f "$C/trainer_state.json" ] && [ -f "$C/adapter_model.safetensors" ] && [ -f "$C/optimizer.pt" ]; do sleep 10; done | |
| sleep 30 | |
| tmux kill-session -t rl4b | |
| echo "{\"time\": \"$(date -u +%FT%TZ)\", \"event\": \"stopped_at\", \"stage\": \"rl_grpo step 40 (lr 1e-5)\"}" >> ../../logs/v2_timeline.jsonl | |
| sleep 20 | |
| echo "{\"time\": \"$(date -u +%FT%TZ)\", \"event\": \"start\", \"stage\": \"rl_grpo_b\"}" >> ../../logs/v2_timeline.jsonl | |
| source .venv/bin/activate | |
| python -u rl_grpo.py --model checkpoints/v2_joint_merged --init-adapter $C --exclude-seen checkpoints/rl_grpo/completions \ | |
| --select data/rl_prompts.jsonl --fresh-pool data/rl_fresh_pool.jsonl --lr 2.5e-5 --steps 20 --seed 19 \ | |
| --out checkpoints/rl_grpo_b > ../../logs/rl_grpo_b.log 2>&1 | |
| echo "RL_EXIT $?" >> ../../logs/rl_grpo_b.log | |
| echo "{\"time\": \"$(date -u +%FT%TZ)\", \"event\": \"done\", \"stage\": \"rl_grpo_b\"}" >> ../../logs/v2_timeline.jsonl | |
| CKPT=checkpoints/rl_grpo_b/final NAME=rl60 bash run_rl_eval.sh > ../../logs/rl_eval_driver.log 2>&1 | |
| echo "DRIVER_EXIT $?" >> ../../logs/rl_eval_driver.log | |
Xet Storage Details
- Size:
- 1.36 kB
- Xet hash:
- b7f9d76e6f8023f66aaed61e6b47ae165cb0a8da9710a864fbcc014156b16288
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.