Merge pull request #7 from intsystems/meshkovvl

add some configs

Author: VseMeshkov

configs/llama_enwik8.yaml

@@ -0,0 +1,41 @@
+# Model configuration
+model:
+  name: "openlm-research/open_llama_3b"
+  # Options:
+  # - "openlm-research/open_llama_3b"
+  # - "Qwen/Qwen2.5-3B"
+  # - "MultiTrickFox/bloom-2b5_Zen"
+  # - "bigscience/bloom-7b1"
+  # - "bigscience/bloom-560m"
+  # - "bigscience/bloom-1b1"
+  
+  torch_dtype: "bfloat16"  # bfloat16, float16, float32
+  device_map: "auto"
+  cache_dir: cache
+  token: null  # HuggingFace token if needed
+  use_bloom: false  # true for BLOOM models, false for LLaMA/Qwen
+
+# Dataset configuration
+dataset:
+  name: "other"
+  subset: "enwik8"
+  split: "train"
+  cache_dir: cache
+  max_length: 512
+  
+# Processing configuration
+processing:
+  batch_size: 96
+  num_workers: 4
+  shuffle: false
+
+# Output configuration
+output:
+  base_dir: "activation_saved/activations_enwik"
+  model_subdir: "open_llama_3b"  # Subdirectory name for this model
+  save_dtype: "float16"  # float16, float32
+  
+# Activation aggregation
+aggregation:
+  method: ["mean", 'last']  # mean, sum, last, none
+  use_attention_mask: true

configs/open_llama_3b_mrpc_classif.yaml

@@ -0,0 +1,9 @@
+path_to_activations: activation_saved/activations_mrpc/open_llama_3b
+path_to_save: all_stats/classifier_stats/open_llama_3b_mrpc_classif.yaml
+target_col: label
+dataset:
+  name: "glue"
+  subset: "mrpc"
+  split: "train"
+  cache_dir: "cache"
+  max_length: 512