diff options
| author | YurenHao0426 <blackhao0426@gmail.com> | 2026-06-13 12:35:36 -0500 |
|---|---|---|
| committer | YurenHao0426 <blackhao0426@gmail.com> | 2026-06-13 12:35:36 -0500 |
| commit | 66e0d8b9fd4d0f7a2231d689c055e26fdf1cf04a (patch) | |
| tree | c29cba61124018755a19b02c9d33e3ad5f2e05cc /research/flossing/run_step6_prefloss.sh | |
Curated export for clone-and-run Maze training (2x A6000) + diagnostics.
trm/hrm pretrain.py carry trajectory-augmentation code (backward-compatible).
Heavy artifacts (checkpoints/wandb/npz) gitignored; see PROVENANCE.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Diffstat (limited to 'research/flossing/run_step6_prefloss.sh')
| -rwxr-xr-x | research/flossing/run_step6_prefloss.sh | 40 |
1 files changed, 40 insertions, 0 deletions
diff --git a/research/flossing/run_step6_prefloss.sh b/research/flossing/run_step6_prefloss.sh new file mode 100755 index 0000000..172bd3e --- /dev/null +++ b/research/flossing/run_step6_prefloss.sh @@ -0,0 +1,40 @@ +#!/bin/bash +# Step 6: Preflossing experiments from step_13020 (epoch 10K, chaos onset) +# Baseline = original HRM training curve (already have 10 ckpts) +# Two conditions: +# B: Engelken prefloss (500 steps Σλ_i²) → task training 3000 steps +# C: CF prefloss (500 steps Σmax(0,λ_i)²) → task training 3000 steps +set -e +cd /home/yurenh2/rrm/research/flossing +source /home/yurenh2/miniconda3/etc/profile.d/conda.sh +conda activate rrm +export CUDA_VISIBLE_DEVICES=0 + +CKPT_ROOT="/home/yurenh2/rrm/hrm/checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 righteous-python" +SRC_CKPT="step_13020" + +echo "[$(date '+%H:%M:%S')] Starting step6 prefloss experiments" >> step6_runner.log + +# B: Engelken prefloss (Σλ_i², two-sided) +echo "[$(date '+%H:%M:%S')] Starting B (engelken prefloss)" >> step6_runner.log +python step6_prefloss.py \ + --ckpt-root "$CKPT_ROOT" --ckpt-name "$SRC_CKPT" \ + --prefloss-steps 500 --train-steps 3000 \ + --floss-mode engelken --floss-lr 1e-4 --train-lr 1e-5 \ + --batch-size 8 --k-lyap 2 --lyap-act-steps 16 \ + --eval-every 100 --eval-n 512 --eval-batch-size 32 \ + --out step6_B_engelken.json > step6_B.log 2>&1 +echo "[$(date '+%H:%M:%S')] B done" >> step6_runner.log + +# C: CF prefloss (Σmax(0,λ_i)², one-sided hinge) +echo "[$(date '+%H:%M:%S')] Starting C (CF prefloss)" >> step6_runner.log +python step6_prefloss.py \ + --ckpt-root "$CKPT_ROOT" --ckpt-name "$SRC_CKPT" \ + --prefloss-steps 500 --train-steps 3000 \ + --floss-mode cf --floss-lr 1e-4 --train-lr 1e-5 \ + --batch-size 8 --k-lyap 2 --lyap-act-steps 16 \ + --eval-every 100 --eval-n 512 --eval-batch-size 32 \ + --out step6_C_cf.json > step6_C.log 2>&1 +echo "[$(date '+%H:%M:%S')] C done" >> step6_runner.log + +echo "[$(date '+%H:%M:%S')] All step6 prefloss experiments complete" >> step6_runner.log |
