Upload all_config.yaml with huggingface_hub
Browse files- all_config.yaml +35 -0
all_config.yaml
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
arch:
|
| 2 |
+
H_cycles: 2
|
| 3 |
+
H_layers: 4
|
| 4 |
+
L_cycles: 2
|
| 5 |
+
L_layers: 4
|
| 6 |
+
expansion: 4
|
| 7 |
+
halt_exploration_prob: 0.1
|
| 8 |
+
halt_max_steps: 16
|
| 9 |
+
hidden_size: 512
|
| 10 |
+
loss:
|
| 11 |
+
loss_type: stablemax_cross_entropy
|
| 12 |
+
name: losses@ACTLossHead
|
| 13 |
+
name: hrm.hrm_act_v1@HierarchicalReasoningModel_ACTV1
|
| 14 |
+
num_heads: 8
|
| 15 |
+
pos_encodings: rope
|
| 16 |
+
puzzle_emb_ndim: 512
|
| 17 |
+
beta1: 0.9
|
| 18 |
+
beta2: 0.95
|
| 19 |
+
checkpoint_every_eval: true
|
| 20 |
+
checkpoint_path: checkpoints/Hitori-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1
|
| 21 |
+
warm-potoo
|
| 22 |
+
data_path: data/hitori-1k-aug-1000
|
| 23 |
+
epochs: 20000
|
| 24 |
+
eval_interval: 2000
|
| 25 |
+
eval_save_outputs: []
|
| 26 |
+
global_batch_size: 384
|
| 27 |
+
lr: 7.0e-05
|
| 28 |
+
lr_min_ratio: 1.0
|
| 29 |
+
lr_warmup_steps: 2000
|
| 30 |
+
project_name: Hitori-1k-aug-1000 ACT-torch
|
| 31 |
+
puzzle_emb_lr: 7.0e-05
|
| 32 |
+
puzzle_emb_weight_decay: 1.0
|
| 33 |
+
run_name: HierarchicalReasoningModel_ACTV1 warm-potoo
|
| 34 |
+
seed: 0
|
| 35 |
+
weight_decay: 1.0
|