rrm workspace: TRM/HRM/SRM code, Maze dataset, dynamical-analysis pipelineHEADmain

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>

1 files changed, 43 insertions, 0 deletions

diff --git a/research/flossing/launch_trm_directional_multi4_long.sh b/research/flossing/launch_trm_directional_multi4_long.sh
new file mode 100755
index 0000000..1a56bcc
--- /dev/null
+++ b/research/flossing/launch_trm_directional_multi4_long.sh
@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+GPU="${1:-3}"
+
+source /home/yurenh2/miniconda3/etc/profile.d/conda.sh
+conda activate rrm
+
+cd /home/yurenh2/rrm/trm
+export WANDB_MODE=offline
+export CUDA_VISIBLE_DEVICES="${GPU}"
+
+python pretrain.py \
+  arch=trm \
+  data_paths='[/home/yurenh2/rrm/data/sudoku-extreme-1k-aug-1000]' \
+  data_paths_test='[]' \
+  evaluators='[]' \
+  +project_name='Sudoku-extreme-1k-aug-1000-ACT-torch' \
+  +run_name='pretrain_mlp_t_sudoku_directional_multi4_parallel_c4_eps003_sigma003' \
+  +checkpoint_path='checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_directional_multi4_parallel_c4_eps003_sigma003' \
+  +load_checkpoint=null \
+  epochs=12500 eval_interval=1250 min_eval_interval=0 checkpoint_every_eval=true \
+  global_batch_size=192 \
+  lr=0.0001 lr_min_ratio=1.0 lr_warmup_steps=2000 \
+  beta1=0.9 beta2=0.95 weight_decay=1.0 \
+  puzzle_emb_lr=0.0001 puzzle_emb_weight_decay=1.0 \
+  ema=true ema_rate=0.999 freeze_weights=false \
+  arch.mlp_t=true arch.pos_encodings=none arch.puzzle_emb_len=16 arch.no_ACT_continue=true \
+  +trajectory_augment=true \
+  +trajectory_parallel=true \
+  +trajectory_n=4 \
+  +trajectory_noise_std=0.03 \
+  +trajectory_noise_min=0.003 \
+  +trajectory_noise_max=0.03 \
+  +trajectory_noise_sampling=loguniform \
+  +trajectory_sigma_start=0.0 \
+  +trajectory_sigma_ramp_steps=5000 \
+  +trajectory_perturb=both \
+  +trajectory_directional=true \
+  +trajectory_directional_candidates=4 \
+  +trajectory_directional_fd_eps=0.03 \
+  +trajectory_directional_horizon=16 \
+  +trajectory_directional_sign=alternate