Replace Iterblock with LLMBlock

scripts/test_freeform_skills.py

@@ -49,7 +49,7 @@
 
 ds = Dataset.from_list(samples)
 
-skills_flow = SynthSkillsFlow(client, teacher_model).get_flow()
+skills_flow = SynthSkillsFlow(client, "mixtral", teacher_model, 1).get_flow()
 skills_pipe = Pipeline(skills_flow)
 
 sdg = SDG([skills_pipe])

scripts/test_grounded_skills.py

@@ -0,0 +1,107 @@
+# Third Party
+from datasets import Dataset
+from openai import OpenAI
+
+# First Party
+from src.instructlab.sdg import SDG
+from src.instructlab.sdg.default_flows import SynthGroundedSkillsFlow
+from src.instructlab.sdg.pipeline import Pipeline
+
+# for vLLM endpoints, the api_key remains "EMPTY"
+openai_api_key = "EMPTY"
+openai_api_base = "Add model endpoint here"
+
+
+client = OpenAI(
+    api_key=openai_api_key,
+    base_url=openai_api_base,
+)
+
+models = client.models.list()
+teacher_model = models.data[0].id
+
+samples = [
+    {
+        "seed_context": """*Ms. Thompson:* Good morning, everyone. Today, we''re here to discuss
+    our customer journey mapping and analysis. I believe this is crucial to understanding
+    our customers'' experiences and improving our services.
+
+
+    *Mr. Patel:* I agree, Lisa. We should start by identifying all touchpoints in
+    our customer journey, from initial contact to post-sale support.
+
+
+    *Ms. Rodriguez:* Yes, and let''s not forget about the emotional aspect of the
+    journey. How do our customers feel at each stage? What are their pain points?
+
+
+    *Mr. Kim:* We can use data from our CRM system to track the customer journey and
+    gather insights. This will help us create a detailed, data-driven map.
+
+
+    *Ms. Johnson:* Once we have the map, we can analyze it to identify areas for improvement.
+    Perhaps there are steps where customers drop off or express dissatisfaction.
+
+
+    *Mr. Davis:* We should also consider the customer''s perspective. Conducting interviews
+    or surveys can provide valuable insights into their thoughts and feelings.
+
+
+    *Ms. Brown:* Absolutely. And once we''ve identified areas for improvement, we
+    can develop strategies to address them. This might involve redesigning certain
+    touchpoints, enhancing our communication, or streamlining processes.
+
+
+    *Mr. Smith:* And we must remember to measure the impact of any changes we make.
+    By tracking key performance indicators, we can determine whether our efforts are
+    successful.
+
+
+    *Ms. Thompson:* Great suggestions, everyone. Let''s divide into teams to tackle
+    different aspects of the customer journey. Team 1 will focus on pre-sale touchpoints,
+    Team 2 on sales, and Team 3 on post-sale support.
+
+
+    *Mr. Patel:* Sounds good. We''ll reconvene in a week to share our findings and
+    discuss next steps.""",
+        "seed_question": """Generate a company wide email based on the given meeting transcript""",
+        "task_description": "Writing emails based on minutes of meeting",
+        "seed_response": """Subject: Recap and Next Steps - Customer Journey Mapping and Analysis\n\
+    \nDear [Company Name] Team,\n\nI hope this email finds you well. Yesterday, we\
+    \ gathered to discuss our customer journey mapping and analysis, aiming to understand\
+    \ our customers' experiences and identify opportunities for improvement. The discussions\
+    \ were fruitful, and I want to share a brief overview of the key points and outcomes.\n\
+    \n1. **Identifying Touchpoints:**\n   Mr. Patel suggested mapping all touchpoints\
+    \ in our customer journey, from initial contact to post-sale support.\n\n2. **Emotional\
+    \ Aspect and Pain Points:**\n   Ms. Rodriguez emphasized the importance of considering\
+    \ the emotional aspect of the journey and identifying customers' pain points at\
+    \ each stage.\n\n3. **Data-Driven Mapping:**\n   Mr. Kim proposed using data from\
+    \ our CRM system to create a detailed, data-driven customer journey map.\n\n4.\
+    \ **Customer Perspective:**\n   Ms. Johnson recommended gathering insights from\
+    \ the customer's perspective through interviews or surveys.\n\n5. **Analysis and\
+    \ Improvement:**\n   Ms. Brown suggested analyzing the customer journey map to\
+    \ identify areas for improvement and developing strategies to address them.\n\n\
+    6. **Measuring Impact:**\n   Mr. Smith stressed the need to measure the impact\
+    \ of any changes made by tracking key performance indicators.\n\nTo facilitate\
+    \ a comprehensive analysis, we have divided into teams to tackle different aspects\
+    \ of the customer journey:\n\n* Team 1: Pre-sale touchpoints\n* Team 2: Sales\n\
+    * Team 3: Post-sale support\n\nEach team will share their findings and discuss\
+    \ next steps in a week.\n\nYour engagement and insights have been invaluable in\
+    \ understanding our customers' experiences and identifying opportunities for improvement.\
+    \ I look forward to our continued collaboration as we work towards enhancing our\
+    \ services and delivering exceptional customer experiences.\n\nBest regards,\n\
+    \n[Your Full Name]\n[Your Position]\n[Company Name]""",
+    }
+]
+
+
+ds = Dataset.from_list(samples)
+
+skills_flow = SynthGroundedSkillsFlow(client, "mixtral", teacher_model, 10).get_flow()
+skills_pipe = Pipeline(skills_flow)
+
+sdg = SDG([skills_pipe])
+gen_data = sdg.generate(ds)
+
+print(gen_data)
+print(gen_data[0])

scripts/test_knowledge.py

@@ -38,8 +38,8 @@
 
 ds = Dataset.from_list(samples)
 
-mmlu_flow = MMLUBenchFlow(client, teacher_model).get_flow()
-knowledge_flow = SynthKnowledgeFlow(client, teacher_model).get_flow()
+mmlu_flow = MMLUBenchFlow(client, "mixtral", teacher_model, 1).get_flow()
+knowledge_flow = SynthKnowledgeFlow(client, "mixtral", teacher_model, 1).get_flow()
 knowledge_pipe = Pipeline(knowledge_flow)
 mmlu_pipe = Pipeline(mmlu_flow)
 

src/instructlab/sdg/configs/skills/evaluate_grounded_pair.yaml

@@ -31,6 +31,7 @@ examples: |
    [End of Score]
 
 generation: |
+  Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, you must rate the answer on a scale of 1 to 3 as mentioned above. 
   Here's the context, question and the answer you need to evaluate:
 
   [Start of Context]
@@ -45,7 +46,6 @@ generation: |
   {answer}
   [End of Answer]
 
-  Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, you must rate the answer on a scale of 1 to 3 as mentioned above. 
   * Return the evaluation between [Start of Evaluation] and [End of Evaluation] tags.
   * Return the score between [Start of Score] and [End of Score] tags.
 

src/instructlab/sdg/configs/skills/evaluate_grounded_questions.yaml

@@ -34,7 +34,7 @@ examples: |
    [End of Score]
 
 generation: |   
-   Here's the context and question you need to evaluate:
+   Here's the context and question you need to evaluate. Return the evaluation between [Start of Evaluation] and [End of Evaluation] tags.
 
    [Start of Context]
    {context}

src/instructlab/sdg/configs/skills/freeform_responses.yaml

@@ -21,13 +21,13 @@ examples: |
   [End of Response]
 
 generation: |
-  Now generate a response to the following prompt. 
+  Now generate a response to the following prompt. Remember to use the same style and format as the example above. 
 
   [Start of Question]
   {question}
   [End of Question]
 
-  Remember to use the same style and format as the example above. Return the response between [Start of Response] and [End of Response] tags.
+  Return the response between [Start of Response] and [End of Response] tags.
 
 start_tags: ["[Start of Response]"]
 end_tags: ["[End of Response]"]

src/instructlab/sdg/configs/skills/grounded_responses.yaml

@@ -26,7 +26,8 @@ examples: |
   [End of Response]
 
 generation: |
-  Now generate a response to the following prompt. Remember to use the same style and format as the example above. Return the response between [Start of Response] and [End of Response] tags.
+  Now generate a response to the following prompt. Remember to use the same style and format as the example above. 
+  Return the response between [Start of Response] and [End of Response] tags.
 
   [Start of Context]
   {context}
@@ -35,6 +36,8 @@ generation: |
   {question}
   [End of Question]
 
+  Return the response between [Start of Response] and [End of Response] tags.
+
 
 start_tags: ["[Start of Response]"]
 end_tags: ["[End of Response]"]