upload

config/arch/hrm.yaml

@@ -0,0 +1,24 @@
+name: recursive_reasoning.hrm@HierarchicalReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 2
+L_cycles: 2
+
+H_layers: 4
+L_layers: 4
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer

config/arch/transformers_baseline.yaml

@@ -0,0 +1,18 @@
+name: recursive_reasoning.transformers_baseline@Model_ACTV2
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 1  # kept for compatibility
+H_layers: 8 
+
+hidden_size: 512  
+num_heads: 12
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope

config/arch/trm.yaml

@@ -0,0 +1,26 @@
+name: recursive_reasoning.trm@TinyRecursiveReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 3
+L_cycles: 6
+
+H_layers: 0
+L_layers: 2
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer
+puzzle_emb_len: 16 # if non-zero, its specified to this value
+no_ACT_continue: True # No continue ACT loss, only use the sigmoid of the halt which makes much more sense

config/arch/trm_hier6.yaml

@@ -0,0 +1,26 @@
+name: recursive_reasoning.trm_hier6@TinyRecursiveReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 3
+L_cycles: 6
+
+H_layers: 0
+L_layers: 2
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer
+puzzle_emb_len: 16 # if non-zero, its specified to this value
+no_ACT_continue: True # No continue ACT loss, only use the sigmoid of the halt which makes much more sense

config/arch/trm_singlez.yaml

@@ -0,0 +1,26 @@
+name: recursive_reasoning.trm_singlez@TinyRecursiveReasoningModel_ACTV1
+loss:
+  name: losses@ACTLossHead
+  loss_type: stablemax_cross_entropy
+
+halt_exploration_prob: 0.1
+halt_max_steps: 16
+
+H_cycles: 3
+L_cycles: 6
+
+H_layers: 0
+L_layers: 2
+
+hidden_size: 512
+num_heads: 8  # min(2, hidden_size // 64)
+expansion: 4
+
+puzzle_emb_ndim: ${.hidden_size}
+
+pos_encodings: rope
+forward_dtype: bfloat16
+
+mlp_t: False # use mlp on L instead of transformer
+puzzle_emb_len: 16 # if non-zero, its specified to this value
+no_ACT_continue: True # No continue ACT loss, only use the sigmoid of the halt which makes much more sense

config/cfg_pretrain.yaml

@@ -0,0 +1,42 @@
+# ARC training config
+
+defaults:
+  - arch: trm
+  - _self_
+
+hydra:
+  output_subdir: null
+
+# Data path
+data_paths: ['data/arc-aug-1000']
+data_paths_test: []
+
+evaluators:
+  - name: arc@ARC
+
+# 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
+
+seed: 0
+min_eval_interval: 0 # when to start the eval
+
+ema: False # use Exponential-Moving-Average
+ema_rate: 0.999 # EMA-rate
+freeze_weights: False # If True, freeze weights and only learn the embeddings