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, 106 insertions, 0 deletions

diff --git a/research/flossing/smoke_step9_hrm_parallel_sampling.json b/research/flossing/smoke_step9_hrm_parallel_sampling.json
new file mode 100644
index 0000000..9d400b4
--- /dev/null
+++ b/research/flossing/smoke_step9_hrm_parallel_sampling.json
@@ -0,0 +1,106 @@
+{
+  "args": {
+    "model": "hrm",
+    "ckpt_root": "/home/yurenh2/rrm/hrm/checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 righteous-python",
+    "ckpt_name": "step_26040",
+    "mode": "multi_perturbed_ce",
+    "train_steps": 2,
+    "batch_size": 2,
+    "lr": 1e-05,
+    "noise_std": 0.001,
+    "noise_min": 0.0001,
+    "noise_max": null,
+    "noise_sampling": "loguniform",
+    "clean_prob": 0.0,
+    "sigma_start": 0.0,
+    "sigma_ramp_steps": 10,
+    "n_trajectories": 4,
+    "rollout_impl": "parallel_fixed",
+    "perturb": "both",
+    "seed": 123,
+    "eval_every": 1,
+    "eval_n": 16,
+    "eval_batch_size": 8,
+    "out": "smoke_step9_hrm_parallel_sampling.json",
+    "save_dir": null,
+    "save_best": true,
+    "save_final": true,
+    "save_every_eval": false
+  },
+  "initial_acc": 0.4375,
+  "initial_tok_acc": 0.816358024691358,
+  "steps": [
+    {
+      "train_step": 1,
+      "loss": 0.2585001587867737,
+      "clean_loss": 0.22608037292957306,
+      "noisy_loss_mean": 0.2693067789077759,
+      "noise_std_target": 0.0001,
+      "noise_std_mean": 9.99999901978299e-05,
+      "noise_std_max": 9.99999901978299e-05,
+      "effective_batch": 8
+    },
+    {
+      "train_step": 2,
+      "loss": 0.08793824911117554,
+      "clean_loss": 0.08808630704879761,
+      "noisy_loss_mean": 0.08788891136646271,
+      "noise_std_target": 0.0002,
+      "noise_std_mean": 0.00015660555800423026,
+      "noise_std_max": 0.00019374192925170064,
+      "effective_batch": 8
+    }
+  ],
+  "evals": [
+    {
+      "kind": "initial",
+      "train_step": 0,
+      "acc": 0.4375,
+      "tok_acc": 0.816358024691358
+    },
+    {
+      "kind": "task",
+      "train_step": 1,
+      "acc": 0.4375,
+      "tok_acc": 0.8240740740740741
+    },
+    {
+      "kind": "task",
+      "train_step": 2,
+      "acc": 0.4375,
+      "tok_acc": 0.8263888888888888
+    },
+    {
+      "kind": "final",
+      "train_step": 2,
+      "acc": 0.4375,
+      "tok_acc": 0.8263888888888888
+    }
+  ],
+  "checkpoints": [
+    {
+      "kind": "best",
+      "train_step": 1,
+      "acc": 0.4375,
+      "path": "smoke_step9_hrm_parallel_sampling_ckpts/best.pt"
+    },
+    {
+      "kind": "best",
+      "train_step": 2,
+      "acc": 0.4375,
+      "path": "smoke_step9_hrm_parallel_sampling_ckpts/best.pt"
+    },
+    {
+      "kind": "final",
+      "train_step": 2,
+      "acc": 0.4375,
+      "path": "smoke_step9_hrm_parallel_sampling_ckpts/final.pt"
+    }
+  ],
+  "best_acc": 0.4375,
+  "best_step": 2,
+  "best_checkpoint": "smoke_step9_hrm_parallel_sampling_ckpts/best.pt",
+  "final_acc": 0.4375,
+  "final_tok_acc": 0.8263888888888888,
+  "final_checkpoint": "smoke_step9_hrm_parallel_sampling_ckpts/final.pt"
+}
\ No newline at end of file