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, 90 insertions, 0 deletions

diff --git a/research/flossing/engelken_paper_faithful/smoke.json b/research/flossing/engelken_paper_faithful/smoke.json
new file mode 100644
index 0000000..ef5cfd9
--- /dev/null
+++ b/research/flossing/engelken_paper_faithful/smoke.json
@@ -0,0 +1,90 @@
+{
+  "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": 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": "research/flossing/engelken_paper_faithful/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": "after_floss",
+      "train_step": 0,
+      "episode": 0,
+      "acc": 0.0,
+      "tok_acc": 0.11265432098765432
+    },
+    {
+      "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