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>

6 files changed, 225 insertions, 0 deletions

diff --git a/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_baseline_nofloss_b8_10000.cmd.sh b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_baseline_nofloss_b8_10000.cmd.sh
new file mode 100755
index 0000000..98411d8
--- /dev/null
+++ b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_baseline_nofloss_b8_10000.cmd.sh
@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "/home/yurenh2/rrm"
+export CUDA_VISIBLE_DEVICES="0"
+export PYTHONUNBUFFERED=1
+exec "/home/yurenh2/miniconda3/envs/rrm/bin/python" research/flossing/step7_interfloss.py \
+  --model trm \
+  --ckpt-root "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_repro" \
+  --ckpt-name __random__ \
+  --init-seed 123 \
+  --train-steps "10000" \
+  --batch-size 8 \
+  --train-lr 1e-4 \
+  --floss-lr 1e-4 \
+  --floss-mode engelken_l2 \
+  --lambda-star 0 \
+  --k-lyap 4 \
+  --lyap-act-steps 4 \
+  --seed 42 \
+  --eval-every 1000 \
+  --eval-n "1000" \
+  --eval-batch-size 64 \
+  --floss-log-every 10 \
+  --train-puzzle-emb \
+  --puzzle-emb-lr 1e-4 \
+  --puzzle-emb-weight-decay 1.0 \
+  --kl-beta 0 \
+  --floss-steps "0" \
+  --interfloss-at "" \
+  --out "/home/yurenh2/rrm/research/flossing/flossing_suite/results/trm_faithful/trm_seed123_baseline_nofloss_b8_10000.json"
diff --git a/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_baseline_nofloss_b8_10000.json b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_baseline_nofloss_b8_10000.json
new file mode 100644
index 0000000..a6f1f2d
--- /dev/null
+++ b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_baseline_nofloss_b8_10000.json
@@ -0,0 +1,45 @@
+{
+  "args": {
+    "model": "trm",
+    "ckpt_root": "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_repro",
+    "ckpt_name": "__random__",
+    "train_steps": 10000,
+    "batch_size": 8,
+    "train_lr": 0.0001,
+    "floss_lr": 0.0001,
+    "floss_steps": 0,
+    "interfloss_at": "",
+    "floss_mode": "engelken_l2",
+    "lambda_star": 0.0,
+    "k_lyap": 4,
+    "lyap_act_steps": 4,
+    "lyap_start_act": 0,
+    "seed": 42,
+    "eval_every": 1000,
+    "eval_n": 1000,
+    "eval_batch_size": 64,
+    "floss_log_every": 10,
+    "eval_after_floss": true,
+    "kl_beta": 0.0,
+    "kl_replay_size": 64,
+    "kl_batch_size": 8,
+    "kl_temperature": 1.0,
+    "init_seed": 123,
+    "train_puzzle_emb": true,
+    "puzzle_emb_lr": 0.0001,
+    "puzzle_emb_weight_decay": 1.0,
+    "out": "/home/yurenh2/rrm/research/flossing/flossing_suite/results/trm_faithful/trm_seed123_baseline_nofloss_b8_10000.json"
+  },
+  "initial_acc": 0.0,
+  "initial_tok_acc": 0.08865432098765431,
+  "task_steps": [],
+  "floss_episodes": [],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.0,
+      "tok_acc": 0.08865432098765431
+    }
+  ]
+}
\ No newline at end of file
diff --git a/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_pre_inter_0_500_b8_k4_10000.cmd.sh b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_pre_inter_0_500_b8_k4_10000.cmd.sh
new file mode 100755
index 0000000..e3f25ab
--- /dev/null
+++ b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_pre_inter_0_500_b8_k4_10000.cmd.sh
@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "/home/yurenh2/rrm"
+export CUDA_VISIBLE_DEVICES="3"
+export PYTHONUNBUFFERED=1
+exec "/home/yurenh2/miniconda3/envs/rrm/bin/python" research/flossing/step7_interfloss.py \
+  --model trm \
+  --ckpt-root "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_repro" \
+  --ckpt-name __random__ \
+  --init-seed 123 \
+  --train-steps "10000" \
+  --batch-size 8 \
+  --train-lr 1e-4 \
+  --floss-lr 1e-4 \
+  --floss-mode engelken_l2 \
+  --lambda-star 0 \
+  --k-lyap 4 \
+  --lyap-act-steps 4 \
+  --seed 42 \
+  --eval-every 1000 \
+  --eval-n "1000" \
+  --eval-batch-size 64 \
+  --floss-log-every 10 \
+  --train-puzzle-emb \
+  --puzzle-emb-lr 1e-4 \
+  --puzzle-emb-weight-decay 1.0 \
+  --kl-beta 0 \
+  --floss-steps "500" \
+  --interfloss-at "0,500" \
+  --out "/home/yurenh2/rrm/research/flossing/flossing_suite/results/trm_faithful/trm_seed123_pre_inter_0_500_b8_k4_10000.json"
diff --git a/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_pre_inter_0_500_b8_k4_10000.json b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_pre_inter_0_500_b8_k4_10000.json
new file mode 100644
index 0000000..6f7ca45
--- /dev/null
+++ b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_pre_inter_0_500_b8_k4_10000.json
@@ -0,0 +1,45 @@
+{
+  "args": {
+    "model": "trm",
+    "ckpt_root": "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_repro",
+    "ckpt_name": "__random__",
+    "train_steps": 10000,
+    "batch_size": 8,
+    "train_lr": 0.0001,
+    "floss_lr": 0.0001,
+    "floss_steps": 500,
+    "interfloss_at": "0,500",
+    "floss_mode": "engelken_l2",
+    "lambda_star": 0.0,
+    "k_lyap": 4,
+    "lyap_act_steps": 4,
+    "lyap_start_act": 0,
+    "seed": 42,
+    "eval_every": 1000,
+    "eval_n": 1000,
+    "eval_batch_size": 64,
+    "floss_log_every": 10,
+    "eval_after_floss": true,
+    "kl_beta": 0.0,
+    "kl_replay_size": 64,
+    "kl_batch_size": 8,
+    "kl_temperature": 1.0,
+    "init_seed": 123,
+    "train_puzzle_emb": true,
+    "puzzle_emb_lr": 0.0001,
+    "puzzle_emb_weight_decay": 1.0,
+    "out": "/home/yurenh2/rrm/research/flossing/flossing_suite/results/trm_faithful/trm_seed123_pre_inter_0_500_b8_k4_10000.json"
+  },
+  "initial_acc": 0.0,
+  "initial_tok_acc": 0.08865432098765431,
+  "task_steps": [],
+  "floss_episodes": [],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.0,
+      "tok_acc": 0.08865432098765431
+    }
+  ]
+}
\ No newline at end of file
diff --git a/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_prefloss_0_b8_k4_10000.cmd.sh b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_prefloss_0_b8_k4_10000.cmd.sh
new file mode 100755
index 0000000..86c445d
--- /dev/null
+++ b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_prefloss_0_b8_k4_10000.cmd.sh
@@ -0,0 +1,30 @@
+#!/usr/bin/env bash
+set -euo pipefail
+cd "/home/yurenh2/rrm"
+export CUDA_VISIBLE_DEVICES="1"
+export PYTHONUNBUFFERED=1
+exec "/home/yurenh2/miniconda3/envs/rrm/bin/python" research/flossing/step7_interfloss.py \
+  --model trm \
+  --ckpt-root "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_repro" \
+  --ckpt-name __random__ \
+  --init-seed 123 \
+  --train-steps "10000" \
+  --batch-size 8 \
+  --train-lr 1e-4 \
+  --floss-lr 1e-4 \
+  --floss-mode engelken_l2 \
+  --lambda-star 0 \
+  --k-lyap 4 \
+  --lyap-act-steps 4 \
+  --seed 42 \
+  --eval-every 1000 \
+  --eval-n "1000" \
+  --eval-batch-size 64 \
+  --floss-log-every 10 \
+  --train-puzzle-emb \
+  --puzzle-emb-lr 1e-4 \
+  --puzzle-emb-weight-decay 1.0 \
+  --kl-beta 0 \
+  --floss-steps "500" \
+  --interfloss-at "0" \
+  --out "/home/yurenh2/rrm/research/flossing/flossing_suite/results/trm_faithful/trm_seed123_prefloss_0_b8_k4_10000.json"
diff --git a/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_prefloss_0_b8_k4_10000.json b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_prefloss_0_b8_k4_10000.json
new file mode 100644
index 0000000..94f4cf6
--- /dev/null
+++ b/research/flossing/flossing_suite/archive_failed_oom_b8/trm_seed123_prefloss_0_b8_k4_10000.json
@@ -0,0 +1,45 @@
+{
+  "args": {
+    "model": "trm",
+    "ckpt_root": "/home/yurenh2/rrm/trm/checkpoints/Sudoku-extreme-1k-aug-1000-ACT-torch/pretrain_mlp_t_sudoku_official_gbs768_repro",
+    "ckpt_name": "__random__",
+    "train_steps": 10000,
+    "batch_size": 8,
+    "train_lr": 0.0001,
+    "floss_lr": 0.0001,
+    "floss_steps": 500,
+    "interfloss_at": "0",
+    "floss_mode": "engelken_l2",
+    "lambda_star": 0.0,
+    "k_lyap": 4,
+    "lyap_act_steps": 4,
+    "lyap_start_act": 0,
+    "seed": 42,
+    "eval_every": 1000,
+    "eval_n": 1000,
+    "eval_batch_size": 64,
+    "floss_log_every": 10,
+    "eval_after_floss": true,
+    "kl_beta": 0.0,
+    "kl_replay_size": 64,
+    "kl_batch_size": 8,
+    "kl_temperature": 1.0,
+    "init_seed": 123,
+    "train_puzzle_emb": true,
+    "puzzle_emb_lr": 0.0001,
+    "puzzle_emb_weight_decay": 1.0,
+    "out": "/home/yurenh2/rrm/research/flossing/flossing_suite/results/trm_faithful/trm_seed123_prefloss_0_b8_k4_10000.json"
+  },
+  "initial_acc": 0.0,
+  "initial_tok_acc": 0.08865432098765431,
+  "task_steps": [],
+  "floss_episodes": [],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.0,
+      "tok_acc": 0.08865432098765431
+    }
+  ]
+}
\ No newline at end of file