Remove the unnecessary SDG class

A pipeline chains together a sequence of blocks, and an SDG chains together
a sequence of pipelines. There is no need for this additional layer - we
can construct a pipeline with the full sequence of blocks, and not chain
pipelines together.

Signed-off-by: Mark McLoughlin <markmc@redhat.com>

Author: markmc

scripts/test_freeform_skills.py

@@ -6,7 +6,6 @@
 from openai import OpenAI
 
 # First Party
-from src.instructlab.sdg import SDG
 from src.instructlab.sdg.pipeline import (
     FULL_PIPELINES_PACKAGE,
     Pipeline,
@@ -60,8 +59,7 @@
 with resources.path(FULL_PIPELINES_PACKAGE, "freeform_skills.yaml") as yaml_path:
     skills_pipe = Pipeline.from_file(ctx, yaml_path)
 
-sdg = SDG([skills_pipe])
-gen_data = sdg.generate(ds)
+gen_data = skills_pipe.generate(ds)
 
 print(gen_data)
 print(gen_data[0])

scripts/test_grounded_skills.py

@@ -6,7 +6,6 @@
 from openai import OpenAI
 
 # First Party
-from src.instructlab.sdg import SDG
 from src.instructlab.sdg.pipeline import (
     FULL_PIPELINES_PACKAGE,
     Pipeline,
@@ -108,8 +107,7 @@
 with resources.path(FULL_PIPELINES_PACKAGE, "grounded_skills.yaml") as yaml_path:
     skills_pipe = Pipeline.from_file(ctx, yaml_path)
 
-sdg = SDG([skills_pipe])
-gen_data = sdg.generate(ds)
+gen_data = skills_pipe.generate(ds)
 
 print(gen_data)
 print(gen_data[0])

scripts/test_knowledge.py

@@ -7,7 +7,6 @@
 from openai import OpenAI
 
 # First Party
-from src.instructlab.sdg import SDG
 from src.instructlab.sdg.pipeline import (
     FULL_PIPELINES_PACKAGE,
     Pipeline,
@@ -47,8 +46,7 @@
 with resources.path(FULL_PIPELINES_PACKAGE, "knowledge.yaml") as yaml_path:
     knowledge_pipe = Pipeline.from_file(ctx, yaml_path)
 
-sdg = SDG([knowledge_pipe])
-mmlubench_data = sdg.generate(ds)
+mmlubench_data = knowledge_pipe.generate(ds)
 
 print(mmlubench_data)
 print(mmlubench_data[0])

src/instructlab/sdg/__init__.py

@@ -21,7 +21,6 @@
     "SamplePopulatorBlock",
     "SelectorBlock",
     "SetToMajorityValueBlock",
-    "SDG",
     "SIMPLE_PIPELINES_PACKAGE",
     "FULL_PIPELINES_PACKAGE",
     "generate_data",
@@ -42,7 +41,6 @@
     PipelineConfigParserError,
     PipelineContext,
 )
-from .sdg import SDG
 from .utilblocks import (
     CombineColumnsBlock,
     DuplicateColumnsBlock,

src/instructlab/sdg/generate_data.py

@@ -30,7 +30,6 @@
     Pipeline,
     PipelineContext,
 )
-from instructlab.sdg.sdg import SDG
 from instructlab.sdg.utils import GenerateException, models
 from instructlab.sdg.utils.taxonomy import (
     leaf_node_to_samples,
@@ -241,9 +240,9 @@ def load_pipeline(yaml_basename):
             return Pipeline.from_file(ctx, os.path.join(pipeline, yaml_basename))
 
     return (
-        SDG([load_pipeline("knowledge.yaml")]),
-        SDG([load_pipeline("freeform_skills.yaml")]),
-        SDG([load_pipeline("grounded_skills.yaml")]),
+        load_pipeline("knowledge.yaml"),
+        load_pipeline("freeform_skills.yaml"),
+        load_pipeline("grounded_skills.yaml"),
     )
 
 
@@ -361,7 +360,9 @@ def generate_data(
         batch_num_workers=num_cpus,
     )
 
-    sdg_knowledge, sdg_freeform_skill, sdg_grounded_skill = _sdg_init(ctx, pipeline)
+    knowledge_pipe, freeform_skills_pipe, grounded_skills_pipe = _sdg_init(
+        ctx, pipeline
+    )
 
     # Make sure checkpointing is disabled (we don't want this pipeline to load checkpoints from the main pipeline)
     mmlu_ctx = dataclasses.replace(ctx, checkpoint_dir=None)
@@ -384,19 +385,19 @@ def generate_data(
             raise GenerateException("Error: No samples found in leaf node.")
 
         if samples[0].get("document"):
-            sdg = sdg_knowledge
+            pipe = knowledge_pipe
             is_knowledge = True
 
         elif samples[0].get("seed_context"):
-            sdg = sdg_grounded_skill
+            pipe = grounded_skills_pipe
 
         else:
-            sdg = sdg_freeform_skill
+            pipe = freeform_skills_pipe
 
         logger.debug("Samples: %s", samples)
         ds = Dataset.from_list(samples)
         logger.debug("Dataset: %s", ds)
-        new_generated_data = sdg.generate(ds)
+        new_generated_data = pipe.generate(ds)
         if len(new_generated_data) == 0:
             raise EmptyDatasetError(
                 "Pipeline stopped: Empty dataset after running pipe"