Automatic dcft datacuration framework upload for oh_v1.3_camel_math_x2

config/Alpaca/alpaca_gpt-4o-mini.yaml

@@ -0,0 +1,78 @@
+operators:  
+- id: load_seed_instructions
+  config:
+    type: function
+    function: data_strategies.Alpaca.utils.load_seed_instructions
+    function_config:
+      seed_tasks_path: dcft/data_strategies/Alpaca/resources/seed_tasks.jsonl
+  
+# NOTE(Ryan): Here we have to guess the number of instructions to generate based on the yeild over time of filtering out samples downstream. 
+- id: create_alpaca_prompts
+  config:
+    type: function
+    function: data_strategies.Alpaca.utils.create_alpaca_prompts
+    function_config:
+      num_instructions_to_generate: 1_000_000
+      num_prompt_instructions: 3
+      output_column: alpaca_prompt
+  input_ids:
+    - load_seed_instructions
+
+- id: generate_instruction_input_output_tuples
+  config:
+    type: completions
+    map: alpaca
+    map_config:
+      alpaca_prompt_column: alpaca_prompt
+      num_seed_instructions: 3 # NOTE(Ryan) used for Alpaca's parsing code, needs to match the number in the previous step
+      output_instruction_column: instruction
+      output_input_column: input
+      output_output_column: output
+    model: gpt-4o-mini
+    batch: False
+  input_ids:
+    - create_alpaca_prompts
+
+- id: filter_instructions_by_heuristics
+  config:
+    type: function
+    function: data_strategies.Alpaca.utils.instructions_heuristics
+    function_config:
+      input_instruction_column: instruction
+      output_filtered_reason_column: filtered_reason
+      output_filtered_decision_column: filtered_decision
+  input_ids:
+    - generate_instruction_input_output_tuples
+
+- id: remove_filtered_instructions
+  config:
+    type: function
+    function: data_strategies.commons.filter_out_truey_values
+    function_config:
+      filter_column: filtered_decision
+  input_ids:
+    - filter_instructions_by_heuristics
+
+# NOTE(Ryan): This is a departure from the original Alpaca recipe which uses ROUGE and 0.7 threshold, we do this to compute similarity scores faster.
+# NOTE(Ryan): We also do not include the seed instructions here, we only check the similarity of the generated instructions against each other.
+- id: filter_instructions_cosine_similarity
+  config:
+    type: function
+    function: data_strategies.Alpaca.utils.filter_instructions_cosine_similarity
+    function_config:
+      similarity_threshold: 0.9
+      input_instruction_column: instruction
+  input_ids:
+    - remove_filtered_instructions
+
+- id: convert_alpaca_to_sharegpt
+  config:
+    type: function
+    function: data_strategies.commons.convert_alpaca_to_sharegpt
+    function_config:
+      input_instruction_column: instruction
+      input_input_column: input
+      input_output_column: output
+      output_sharegpt_column: conversations
+  input_ids:
+    - filter_instructions_cosine_similarity