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>

4 files changed, 361 insertions, 0 deletions

diff --git a/research/flossing/flossing_suite/results/smoke/toy_smoke.json b/research/flossing/flossing_suite/results/smoke/toy_smoke.json
new file mode 100644
index 0000000..6870365
--- /dev/null
+++ b/research/flossing/flossing_suite/results/smoke/toy_smoke.json
@@ -0,0 +1,88 @@
+{
+  "config": {
+    "hidden_size": 16,
+    "train_epochs": 3,
+    "inter_period": 100,
+    "inter_epochs": 0,
+    "pre_epochs": 2,
+    "max_inter_episodes": 0,
+    "batch_size": 4,
+    "input_dim": 1,
+    "train_steps": 40,
+    "lyap_steps": 30,
+    "floss_input_steps": 40,
+    "seed_ic": 1,
+    "seed_input": 1,
+    "seed_net": 1,
+    "seed_ons": 1,
+    "lr": 0.001,
+    "beta1": 0.9,
+    "beta2": 0.999,
+    "init_type": 1,
+    "recurrent_gain": 1.0,
+    "recurrent_mean_gain": 0.0,
+    "input_scale": 1.0,
+    "delay": 10,
+    "ws_std": 1.0,
+    "ws_mean": 0.0,
+    "wr_std": 1.0,
+    "wr_mean": 0.0,
+    "b_std": 0.1,
+    "b_mean": 0.0,
+    "n_lyap": 4,
+    "task": -1,
+    "lyap_target": 0.0,
+    "eval_every": 1,
+    "eval_batches": 1,
+    "log_every_floss": 50,
+    "device": "cuda",
+    "out": "/home/yurenh2/rrm/research/flossing/flossing_suite/results/smoke/toy_smoke.json"
+  },
+  "evals": [
+    {
+      "epoch": 1,
+      "train_loss": 0.8864837288856506,
+      "train_accuracy": 0.4880952380952381,
+      "eval_loss": 0.8243135213851929,
+      "eval_accuracy": 0.5476190476190477,
+      "elapsed": 0.8400521278381348
+    },
+    {
+      "epoch": 2,
+      "train_loss": 0.9001060128211975,
+      "train_accuracy": 0.4880952380952381,
+      "eval_loss": 0.9111667275428772,
+      "eval_accuracy": 0.5238095238095238,
+      "elapsed": 0.8638095855712891
+    },
+    {
+      "epoch": 3,
+      "train_loss": 0.9251764416694641,
+      "train_accuracy": 0.4880952380952381,
+      "eval_loss": 0.7443239688873291,
+      "eval_accuracy": 0.6190476190476191,
+      "elapsed": 0.8863844871520996
+    }
+  ],
+  "floss": [
+    {
+      "epoch": 1,
+      "kind": "pre",
+      "floss_step": 1,
+      "loss": 0.18372014164924622,
+      "lambda_mean": -0.4003114700317383,
+      "lambda_1": -0.22871017456054688,
+      "elapsed": 0.7358992099761963
+    },
+    {
+      "epoch": 1,
+      "kind": "pre",
+      "floss_step": 2,
+      "loss": 0.3812721371650696,
+      "lambda_mean": -0.5845165252685547,
+      "lambda_1": -0.29521623253822327,
+      "elapsed": 0.7633428573608398
+    }
+  ],
+  "task": []
+}
\ No newline at end of file
diff --git a/research/flossing/flossing_suite/results/smoke/trm_step7_decoupled_periodic_smoke.json b/research/flossing/flossing_suite/results/smoke/trm_step7_decoupled_periodic_smoke.json
new file mode 100644
index 0000000..927e492
--- /dev/null
+++ b/research/flossing/flossing_suite/results/smoke/trm_step7_decoupled_periodic_smoke.json
@@ -0,0 +1,120 @@
+{
+  "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": 2,
+    "batch_size": 8,
+    "task_batch_size": 8,
+    "floss_batch_size": 2,
+    "train_lr": 0.0001,
+    "floss_lr": 0.0001,
+    "floss_steps": 1,
+    "interfloss_at": "0",
+    "interfloss_every": 1,
+    "interfloss_start": 1,
+    "interfloss_stop": 1,
+    "floss_mode": "engelken_l2",
+    "lambda_star": 0.0,
+    "k_lyap": 1,
+    "lyap_act_steps": 1,
+    "lyap_start_act": 0,
+    "seed": 42,
+    "eval_every": 1,
+    "eval_n": 8,
+    "eval_batch_size": 8,
+    "floss_log_every": 1,
+    "eval_after_floss": false,
+    "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": "research/flossing/flossing_suite/results/smoke/trm_step7_decoupled_periodic_smoke.json"
+  },
+  "initial_acc": 0.0,
+  "initial_tok_acc": 0.07561728395061729,
+  "interfloss_steps": [
+    0,
+    1
+  ],
+  "task_steps": [
+    {
+      "train_step": 1,
+      "sup_loss": 2.57209548232992
+    },
+    {
+      "train_step": 2,
+      "sup_loss": 2.56865356540629
+    }
+  ],
+  "floss_episodes": [
+    {
+      "episode": 0,
+      "train_step": 0,
+      "steps": [
+        {
+          "step": 0,
+          "loss": 0.04431530460715294,
+          "floss_loss": 0.04431530460715294,
+          "kl_loss": 0.0,
+          "lyap1_mean": 0.21050173044204712,
+          "lyap1_max": 0.2125818431377411,
+          "lyap_mean": 0.21050173044204712,
+          "volume_mean": 0.21050173044204712,
+          "volume_max": 0.2125818431377411,
+          "frac_active": 1.0
+        }
+      ]
+    },
+    {
+      "episode": 1,
+      "train_step": 1,
+      "steps": [
+        {
+          "step": 0,
+          "loss": 0.04459226503968239,
+          "floss_loss": 0.04459226503968239,
+          "kl_loss": 0.0,
+          "lyap1_mean": 0.2111663520336151,
+          "lyap1_max": 0.21218420565128326,
+          "lyap_mean": 0.2111663520336151,
+          "volume_mean": 0.2111663520336151,
+          "volume_max": 0.21218420565128326,
+          "frac_active": 1.0
+        }
+      ]
+    }
+  ],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.0,
+      "tok_acc": 0.07561728395061729
+    },
+    {
+      "kind": "task",
+      "train_step": 1,
+      "acc": 0.0,
+      "tok_acc": 0.07253086419753087
+    },
+    {
+      "kind": "task",
+      "train_step": 2,
+      "acc": 0.0,
+      "tok_acc": 0.08796296296296297
+    },
+    {
+      "kind": "final",
+      "train_step": 2,
+      "acc": 0.0,
+      "tok_acc": 0.08796296296296297
+    }
+  ],
+  "final_acc": 0.0,
+  "final_tok_acc": 0.08796296296296297
+}
\ No newline at end of file
diff --git a/research/flossing/flossing_suite/results/smoke/trm_step7_smoke.json b/research/flossing/flossing_suite/results/smoke/trm_step7_smoke.json
new file mode 100644
index 0000000..bb1b789
--- /dev/null
+++ b/research/flossing/flossing_suite/results/smoke/trm_step7_smoke.json
@@ -0,0 +1,83 @@
+{
+  "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": 1,
+    "batch_size": 2,
+    "train_lr": 0.0001,
+    "floss_lr": 0.0001,
+    "floss_steps": 1,
+    "interfloss_at": "0",
+    "floss_mode": "engelken_l2",
+    "lambda_star": 0.0,
+    "k_lyap": 1,
+    "lyap_act_steps": 1,
+    "lyap_start_act": 0,
+    "seed": 42,
+    "eval_every": 1,
+    "eval_n": 8,
+    "eval_batch_size": 8,
+    "floss_log_every": 1,
+    "eval_after_floss": false,
+    "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/smoke/trm_step7_smoke.json"
+  },
+  "initial_acc": 0.0,
+  "initial_tok_acc": 0.07561728395061729,
+  "task_steps": [
+    {
+      "train_step": 1,
+      "sup_loss": 2.570638557821803
+    }
+  ],
+  "floss_episodes": [
+    {
+      "episode": 0,
+      "train_step": 0,
+      "steps": [
+        {
+          "step": 0,
+          "loss": 0.04431530460715294,
+          "floss_loss": 0.04431530460715294,
+          "kl_loss": 0.0,
+          "lyap1_mean": 0.21050173044204712,
+          "lyap1_max": 0.2125818431377411,
+          "lyap_mean": 0.21050173044204712,
+          "volume_mean": 0.21050173044204712,
+          "volume_max": 0.2125818431377411,
+          "frac_active": 1.0
+        }
+      ]
+    }
+  ],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.0,
+      "tok_acc": 0.07561728395061729
+    },
+    {
+      "kind": "task",
+      "train_step": 1,
+      "acc": 0.0,
+      "tok_acc": 0.08641975308641975
+    },
+    {
+      "kind": "final",
+      "train_step": 1,
+      "acc": 0.0,
+      "tok_acc": 0.08641975308641975
+    }
+  ],
+  "final_acc": 0.0,
+  "final_tok_acc": 0.08641975308641975
+}
\ No newline at end of file
diff --git a/research/flossing/flossing_suite/results/smoke/trm_task_batch32_smoke.json b/research/flossing/flossing_suite/results/smoke/trm_task_batch32_smoke.json
new file mode 100644
index 0000000..c6df53e
--- /dev/null
+++ b/research/flossing/flossing_suite/results/smoke/trm_task_batch32_smoke.json
@@ -0,0 +1,70 @@
+{
+  "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": 1,
+    "batch_size": 32,
+    "task_batch_size": 32,
+    "floss_batch_size": 4,
+    "train_lr": 0.0001,
+    "floss_lr": 0.0001,
+    "floss_steps": 0,
+    "interfloss_at": "",
+    "interfloss_every": 0,
+    "interfloss_start": 0,
+    "interfloss_stop": -1,
+    "floss_mode": "engelken_l2",
+    "lambda_star": 0.0,
+    "k_lyap": 4,
+    "lyap_act_steps": 4,
+    "lyap_start_act": 0,
+    "seed": 43,
+    "eval_every": 1,
+    "eval_n": 8,
+    "eval_batch_size": 8,
+    "floss_log_every": 1,
+    "eval_after_floss": false,
+    "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": "research/flossing/flossing_suite/results/smoke/trm_task_batch32_smoke.json"
+  },
+  "initial_acc": 0.0,
+  "initial_tok_acc": 0.07561728395061729,
+  "interfloss_steps": [],
+  "task_steps": [
+    {
+      "train_step": 1,
+      "sup_loss": 2.581977193130616
+    }
+  ],
+  "floss_episodes": [],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.0,
+      "tok_acc": 0.07561728395061729
+    },
+    {
+      "kind": "task",
+      "train_step": 1,
+      "acc": 0.0,
+      "tok_acc": 0.09722222222222222
+    },
+    {
+      "kind": "final",
+      "train_step": 1,
+      "acc": 0.0,
+      "tok_acc": 0.09722222222222222
+    }
+  ],
+  "final_acc": 0.0,
+  "final_tok_acc": 0.09722222222222222
+}
\ No newline at end of file