From 66e0d8b9fd4d0f7a2231d689c055e26fdf1cf04a Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Sat, 13 Jun 2026 12:35:36 -0500 Subject: rrm workspace: TRM/HRM/SRM code, Maze dataset, dynamical-analysis pipeline 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 --- hrm/config/cfg_pretrain.yaml | 31 +++++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+) create mode 100644 hrm/config/cfg_pretrain.yaml (limited to 'hrm/config/cfg_pretrain.yaml') diff --git a/hrm/config/cfg_pretrain.yaml b/hrm/config/cfg_pretrain.yaml new file mode 100644 index 0000000..51c55a0 --- /dev/null +++ b/hrm/config/cfg_pretrain.yaml @@ -0,0 +1,31 @@ +# ARC training config + +defaults: + - arch: hrm_v1 + - _self_ + +hydra: + output_subdir: null + +# Data path +data_path: data/arc-aug-1000 + +# Hyperparams - Training +global_batch_size: 768 + +epochs: 100000 +eval_interval: 10000 +checkpoint_every_eval: True + +lr: 1e-4 +lr_min_ratio: 1.0 +lr_warmup_steps: 2000 + +# Standard hyperparameter settings for LM, as used in Llama +beta1: 0.9 +beta2: 0.95 +weight_decay: 0.1 +puzzle_emb_weight_decay: 0.1 + +# Hyperparams - Puzzle embeddings training +puzzle_emb_lr: 1e-2 -- cgit v1.2.3