Param Passthrough and Consistent Tutorial Script (stanfordnlp#3)

* Add param passthrough and default banking77 tutorial

* Add more threads

* Update banking tutorial

---------

Co-authored-by: Noah Ziems <nziems2@nziems2@nd.edu>

Author: Ziems

banking77_tutorial.py

@@ -0,0 +1,75 @@
+import dspy
+import random
+from dspy.datasets import DataLoader
+from datasets import load_dataset
+
+# Load the Banking77 dataset.
+CLASSES = load_dataset("PolyAI/banking77", split="train", trust_remote_code=True).features['label'].names
+kwargs = dict(fields=("text", "label"), input_keys=("text",), split="train", trust_remote_code=True)
+
+# Load the first 2000 examples from the dataset, and assign a hint to each *training* example.
+raw_data = [
+    dspy.Example(x, label=CLASSES[x.label]).with_inputs("text")
+    for x in DataLoader().from_huggingface(dataset_name="PolyAI/banking77", **kwargs)[:2000]
+]
+
+random.Random(0).shuffle(raw_data)
+print(len(CLASSES), CLASSES[:10])
+
+trainset = raw_data[:1500] # 1500 examples for training
+valset = raw_data[1500:1600] # 100 examples for validation
+print(trainset[0])
+
+classify = dspy.ChainOfThought(f"text -> label: Literal{CLASSES}")
+
+from dspy.clients.lm_local_arbor import ArborProvider
+port = 7453
+arbor_api_base = f"http://localhost:{port}/v1/"
+api_key = "arbor"
+provider = ArborProvider()
+
+# student_lm_name = "meta-llama/Llama-3.2-3B-Instruct"
+# student_lm_name = "Qwen/Qwen2.5-1.5B-Instruct"
+student_lm_name = "Qwen/Qwen3-0.6B"
+student_lm = dspy.LM(model=f"openai/arbor:{student_lm_name}", provider=provider, temperature=0.5, api_base=arbor_api_base, api_key=api_key, max_tokens=3000)
+
+student_classify = classify.deepcopy()
+student_classify.set_lm(student_lm)
+
+metric = (lambda x, y, trace=None: x.label == y.label)
+
+
+from dspy.teleprompt.grpo import GRPO
+train_kwargs = {
+    "update_interval": 10,
+    "per_device_train_batch_size": 4,
+    "temperature": 0.5,
+    "beta": 0.04,
+    "learning_rate": 1e-5,
+    "gradient_accumulation_steps": 1,
+    "gradient_checkpointing": True,
+    "bf16": True,
+    "lr_scheduler_type": "constant_with_warmup",
+}
+
+compiler = GRPO(
+    metric=metric,
+    multitask=True,
+    num_dspy_examples_per_grpo_step=4,
+    num_samples_per_input=4,
+    exclude_demos=True,
+    num_train_steps=500,
+    num_threads=4,
+    use_train_as_val=False,
+    num_steps_for_val=100,
+    train_kwargs=train_kwargs,
+)
+
+classify_ft = compiler.compile(
+    student=student_classify,
+    trainset=trainset,
+    valset=valset,
+)
+
+# evaluate = dspy.Evaluate(devset=valset, metric=metric, display_progress=True, display_table=5, num_threads=16)
+# print(evaluate(classify_ft))

dspy/clients/lm_local_arbor.py

@@ -44,9 +44,15 @@ def status(self) -> TrainingStatus:
 
 class ArborReinforceJob(ReinforceJob):
     DEFAULT_TRAIN_KWARGS = {
-        "update_interval": 5,
+        "update_interval": 10,
         "temperature": 0.9,
         "beta": 0.04,
+        "per_device_train_batch_size": 8,
+        "learning_rate": 1e-5,
+        "gradient_accumulation_steps": 1,
+        "gradient_checkpointing": True,
+        "bf16": True,
+        "lr_scheduler_type": "constant_with_warmup",
     }
 
     def __init__(self, lm: "LM", train_kwargs: Dict[str, Any]):
@@ -66,6 +72,12 @@ def initialize(self):
         update_interval = self.train_kwargs.get("update_interval", self.DEFAULT_TRAIN_KWARGS["update_interval"])
         temperature = self.train_kwargs.get("temperature", self.DEFAULT_TRAIN_KWARGS["temperature"])
         beta = self.train_kwargs.get("beta", self.DEFAULT_TRAIN_KWARGS["beta"])
+        per_device_train_batch_size = self.train_kwargs.get("per_device_train_batch_size", self.DEFAULT_TRAIN_KWARGS["per_device_train_batch_size"])
+        learning_rate = self.train_kwargs.get("learning_rate", self.DEFAULT_TRAIN_KWARGS["learning_rate"])
+        gradient_accumulation_steps = self.train_kwargs.get("gradient_accumulation_steps", self.DEFAULT_TRAIN_KWARGS["gradient_accumulation_steps"])
+        gradient_checkpointing = self.train_kwargs.get("gradient_checkpointing", self.DEFAULT_TRAIN_KWARGS["gradient_checkpointing"])
+        bf16 = self.train_kwargs.get("bf16", self.DEFAULT_TRAIN_KWARGS["bf16"])
+        lr_scheduler_type = self.train_kwargs.get("lr_scheduler_type", self.DEFAULT_TRAIN_KWARGS["lr_scheduler_type"])
 
         api_base = self.lm.kwargs["api_base"]
         # api_key = self.lm.kwargs["api_key"]
@@ -79,6 +91,12 @@ def initialize(self):
             'update_interval': update_interval,
             'temperature': temperature,
             'beta': beta,
+            'per_device_train_batch_size': per_device_train_batch_size,
+            'learning_rate': learning_rate,
+            'gradient_accumulation_steps': gradient_accumulation_steps,
+            'gradient_checkpointing': gradient_checkpointing,
+            'bf16': bf16,
+            'lr_scheduler_type': lr_scheduler_type,
         }
         url = f"{api_base}fine_tuning/grpo/initialize"
         headers = {'Content-Type': 'application/json'}