feat: making Granite 4 the default model (#178)

* Granite 4 is now default

* Granite 4 is now default

* Granite 4 notebbok updates

Author: HendrikStrobelt

docs/examples/best_of_n/prm.py

@@ -2,6 +2,7 @@
 
 from docs.examples.helper import w
 from mellea import start_session
+from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
 from mellea.backends.process_reward_models.huggingface.prms import (
     HFGenerativePRM,
     HFRegressionPRM,
@@ -11,7 +12,11 @@
 from mellea.stdlib.sampling.best_of_n import BestofNSamplingStrategy
 
 # create a session for the generator using Granite 3.3 8B on Huggingface and a simple context [see below]
-m = start_session(backend_name="hf", model_options={ModelOption.MAX_NEW_TOKENS: 512})
+m = start_session(
+    backend_name="hf",
+    model_id=IBM_GRANITE_3_3_8B,
+    model_options={ModelOption.MAX_NEW_TOKENS: 512},
+)
 
 # initialize the PRM model
 prm_model = HFGenerativePRM(

docs/examples/generative_slots/generative_cot.py renamed to docs/examples/generative_slots/generative_gsm8k.py

@@ -130,11 +130,11 @@ def extract_final_short_answer(
 
 if __name__ == "__main__":
     scores = []
+    m = start_session()
+
     for question, target in (
         x.values() for x in load_dataset("gsm8k", "main", split="train[:100]")
     ):
-        m = start_session()
-
         target = int(target.split("####")[-1])
         response = compute_chain_of_thought_and_final_answer(m, question=question)
         for step in response.step_by_step_solution:

docs/examples/generative_slots/generative_slots.py

@@ -15,8 +15,8 @@ def generate_summary(text: str) -> str:
 
 
 if __name__ == "__main__":
-    with start_session():
-        sentiment_component = classify_sentiment(text="I love this!")
+    with start_session() as m:
+        sentiment_component = classify_sentiment(m, text="I love this!")
         print("Output sentiment is : ", sentiment_component)
 
         summary = generate_summary(

docs/examples/information_extraction/101_with_gen_slots.py

@@ -3,12 +3,14 @@
 from mellea import generative, start_session
 from mellea.backends import model_ids
 
-m = start_session(model_id=model_ids.MISTRALAI_MISTRAL_0_3_7B)
+m = start_session()
 
 
 @generative
 def extract_all_person_names(doc: str) -> list[str]:
-    """Given a document, extract all person names. Return these names as list of strings."""
+    """
+    Given a document, extract names of ALL mentioned persons. Return these names as list of strings.
+    """
 
 
 # ref: https://www.nytimes.com/2012/05/20/world/world-leaders-at-us-meeting-urge-growth-not-austerity.html

docs/examples/information_extraction/advanced_with_m_instruct.py

@@ -35,11 +35,11 @@ def _at_least_(t: str) -> bool:
 
 
 # start session
-m = start_session(model_id=model_ids.MISTRALAI_MISTRAL_0_3_7B)
+m = start_session()
 
 # run extraction using grounding context and sampling strategy
 sampled_p_names = m.instruct(
-    "Extract the person names from the document (doc1).",
+    "Extract ALL person names from the document (doc1).",
     grounding_context={"doc1": NYTimes_text},
     requirements=[check(None, validation_fn=simple_validate(at_least_n(2)))],
     strategy=RejectionSamplingStrategy(loop_budget=5),

docs/examples/instruct_validate_repair/101_email.py

@@ -17,10 +17,11 @@
 # # start_session() is equivalent to:
 # from mellea.backends import model_ids
 # from mellea.backends.ollama import OllamaModelBackend
-# from mellea import MelleaSession, SimpleContext
+# from mellea import MelleaSession
+# from mellea.stdlib.base import SimpleContext
 # m = MelleaSession(
 #     backend=OllamaModelBackend(
-#         model_id=model_ids.IBM_GRANITE_3_3_8B,
+#         model_id=model_ids.IBM_GRANITE_4_MICRO_3B,
 #         model_options={ModelOption.MAX_NEW_TOKENS: 200},
 #     ),
 #     ctx=SimpleContext()

docs/examples/instruct_validate_repair/101_email_comparison.py

@@ -11,7 +11,7 @@
     "Write a very funny email to invite all interns to the office party."
 )
 print(
-    f"***** email 1 ****\n{w(email_v1)}\n*******email 2 ******\n{w(email_v2)}\n*******"
+    f"***** email 1 ****\n{w(email_v1)}\n*******email 2 ******\n{w(email_v2)}\n\n*******"
 )
 
 # Use the emails as grounding context to evaluate which one is quirkier

docs/examples/instruct_validate_repair/101_email_with_requirements.py

@@ -8,7 +8,7 @@
 # write an email with automatic requirement checking.
 email_v1 = m.instruct(
     "Write an email to invite all interns to the office party.",
-    requirements=["be formal", "Use 'Dear interns' as greeting."],
+    requirements=["be formal", "Use 'Dear Interns' as greeting."],
 )
 
 # print result

docs/examples/instruct_validate_repair/101_email_with_validate.py

@@ -1,14 +1,15 @@
 from docs.examples.helper import req_print, w
 from mellea import start_session
+from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
 from mellea.backends.types import ModelOption
 from mellea.stdlib.sampling import RejectionSamplingStrategy
 
-# create a session using Granite 3.3 8B on Ollama and a simple context [see below]
+# create a session using Granite 4 Micro (3B) on Ollama and a simple context [see below]
 m = start_session(model_options={ModelOption.MAX_NEW_TOKENS: 200})
 
 email_v2_samples = m.instruct(
-    "Write an email to invite all interns to the office party.",
-    requirements=["be formal", "Use 'Dear interns' as greeting."],
+    "Write a very short email to invite all interns to the office party.",
+    requirements=["Use formal language.", "Use 'Dear Interns' as greeting."],
     strategy=RejectionSamplingStrategy(loop_budget=3),
     return_sampling_results=True,
 )

docs/examples/mini_researcher/researcher.py

@@ -20,7 +20,7 @@
 @cache
 def get_session():
     """Get M session (change model here)."""
-    return MelleaSession(backend=OllamaModelBackend(model_ids.IBM_GRANITE_3_3_8B))
+    return MelleaSession(backend=OllamaModelBackend(model_ids.IBM_GRANITE_4_MICRO_3B))
 
 
 @cache