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, 68 insertions, 0 deletions

diff --git a/research/flossing/launch_multi4_repro_config.sh b/research/flossing/launch_multi4_repro_config.sh
new file mode 100755
index 0000000..8cb5b25
--- /dev/null
+++ b/research/flossing/launch_multi4_repro_config.sh
@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+CASE="${1:?case required}"
+GPU="${2:-0}"
+
+source /home/yurenh2/miniconda3/etc/profile.d/conda.sh
+conda activate rrm
+
+case "${CASE}" in
+  hrm)
+    cd /home/yurenh2/rrm/hrm
+    export WANDB_MODE=offline
+    export CUDA_VISIBLE_DEVICES="${GPU}"
+    python pretrain.py \
+      data_path=/home/yurenh2/rrm/data/sudoku-extreme-1k-aug-1000 \
+      +project_name='Sudoku-extreme-1k-aug-1000 ACT-torch' \
+      +run_name='HierarchicalReasoningModel_ACTV1 multi4-loguniform-repro' \
+      +checkpoint_path='checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 multi4-loguniform-repro' \
+      epochs=20000 eval_interval=2000 checkpoint_every_eval=true \
+      global_batch_size=768 \
+      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 \
+      +trajectory_augment=true \
+      +trajectory_n=4 \
+      +trajectory_noise_std=0.001 \
+      +trajectory_noise_min=0.00003 \
+      +trajectory_noise_max=0.003 \
+      +trajectory_noise_sampling=loguniform \
+      +trajectory_sigma_start=0.0 \
+      +trajectory_sigma_ramp_steps=5000 \
+      +trajectory_perturb=both
+    ;;
+  trm)
+    cd /home/yurenh2/rrm/trm
+    export WANDB_MODE=offline
+    export CUDA_VISIBLE_DEVICES="${GPU}"
+    python pretrain.py \
+      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_multi4_loguniform_repro' \
+      +checkpoint_path='checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_multi4_loguniform_repro' \
+      +load_checkpoint=null \
+      epochs=50000 eval_interval=5000 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_n=4 \
+      +trajectory_noise_std=0.001 \
+      +trajectory_noise_min=0.00003 \
+      +trajectory_noise_max=0.003 \
+      +trajectory_noise_sampling=loguniform \
+      +trajectory_sigma_start=0.0 \
+      +trajectory_sigma_ramp_steps=5000 \
+      +trajectory_perturb=both
+    ;;
+  *)
+    echo "unknown case: ${CASE}" >&2
+    exit 2
+    ;;
+esac