Dev finetune update (#1698)

* Adapter updates

* Client updates

* Add provider

* Add cache dir utils

* Add BootstrapFinetune

* Add BetterTogether draft

* Prepare PR

* Add AnyScale changes -- ruff

* Teporarily remove BetterTogether

* Remove BetterTogether import

* Add comment

* Replace OpenAI client call with library call

* Add OpenAI models list to check valid models

* Temporarily switch to print

* Prepare BootstrapFinetune for BetterTogether

* Add BetterTogether

* Add dev notebook

* Revamp ChainOfThoughtWithHint and adjust max auto valset of MIPROv2

* fix

* ruff fixes

* disable cot_hint tests

* unsafe ruff fixes

---------

Co-authored-by: Omar Khattab <[email protected]>

Author: dilarasoylu

dsp/utils/settings_v2.py

@@ -73,7 +73,7 @@ def main():
             futures = {executor.submit(thread_wrapper, sample_program, parent_tid, arg) for arg in range(3)}
 
             for future in as_completed(futures):
-                res = future.result()
+                future.result()
 
         print(f"Main thread {parent_tid} config after threads: {dsp_settings._get_current_config()}")
 

dspy/__init__.py

@@ -67,6 +67,8 @@
 BootstrapFewShot = dspy.teleprompt.BootstrapFewShot
 BootstrapFewShotWithRandomSearch = dspy.teleprompt.BootstrapFewShotWithRandomSearch
 BootstrapRS = dspy.teleprompt.BootstrapFewShotWithRandomSearch
+BootstrapFinetune = dspy.teleprompt.BootstrapFinetune
+BetterTogether = dspy.teleprompt.BetterTogether
 COPRO = dspy.teleprompt.COPRO
 MIPROv2 = dspy.teleprompt.MIPROv2
 Ensemble = dspy.teleprompt.Ensemble

dspy/adapters/base.py

@@ -1,19 +1,8 @@
-import abc
-from dspy.utils.callback import with_callbacks
-
-class Adapter:
-    @abc.abstractmethod
-    def format(self, signature, demos, inputs):
-        """
-        Format the input data for the LLM.
-        """
+from abc import ABC, abstractmethod
 
-    @abc.abstractmethod
-    def parse(self, signature, completion):
-        """
-        Parse the output data from the LLM.
-        """
+from dspy.utils.callback import with_callbacks
 
+class Adapter(ABC):
     def __init__(self, callbacks=None):
         self.callbacks = callbacks or []
 
@@ -31,7 +20,6 @@ def __call__(self, lm, lm_kwargs, signature, demos, inputs, _parse_values=True):
         outputs = lm(**inputs_, **lm_kwargs)
         values = []
 
-        
         try:
             for output in outputs:
                 value = self.parse(signature, output, _parse_values=_parse_values)
@@ -45,3 +33,13 @@ def __call__(self, lm, lm_kwargs, signature, demos, inputs, _parse_values=True):
                 return JSONAdapter()(lm, lm_kwargs, signature, demos, inputs, _parse_values=_parse_values)
             raise e
 
+    @abstractmethod
+    def format(self, signature, demos, inputs):
+       raise NotImplementedError
+
+    @abstractmethod
+    def parse(self, signature, completion, _parse_values):
+       raise NotImplementedError
+
+    def format_finetune_data(self, signature, demos, inputs, outputs):
+        raise NotImplementedError

dspy/adapters/chat_adapter.py

@@ -34,10 +34,6 @@ class FieldInfoWithName(NamedTuple):
 BuiltInCompletedOutputFieldInfo = FieldInfoWithName(name="completed", info=OutputField())
 
 class ChatAdapter(Adapter):
-    """
-    ChatAdapter is used to format and parse data for chat-based LLMs.
-    """
-
     def format(self, signature: Signature, demos: list[dict[str, Any]], inputs: dict[str, Any]) -> list[dict[str, Any]]:
         messages: list[dict[str, Any]] = []
 
@@ -90,6 +86,19 @@ def parse(self, signature, completion, _parse_values=True):
 
         return fields
 
+    # TODO(PR): Looks ok?
+    def format_finetune_data(self, signature, demos, inputs, outputs):
+        # Get system + user messages
+        messages = self.format(signature, demos, inputs)
+
+        # Add the assistant message
+        role = "assistant"
+        incomplete = False
+        assistant_message = format_turn(signature, outputs, role, incomplete)
+        messages.append(assistant_message)
+
+        # Wrap the messages in a dictionary with a "messages" key
+        return dict(messages=messages)
     def format_turn(self, signature, values, role, incomplete=False):
         return format_turn(signature, values, role, incomplete)
 

dspy/clients/__init__.py

@@ -1,4 +1,5 @@
 from .lm import LM
+from .provider import Provider, TrainingJob
 from .base_lm import BaseLM, inspect_history
 from .embedding import Embedding
 import litellm

dspy/clients/anyscale.py

@@ -7,7 +7,7 @@
 
 from dspy.clients.finetune import (
     FinetuneJob,
-    TrainingMethod,
+    # TrainingMethod,
     save_data,
 )
 from dspy.clients.openai import openai_data_validation
@@ -32,7 +32,7 @@
 def is_anyscale_model(model: str) -> bool:
     """Check if the model is an AnyScale model."""
     # TODO: This needs to be implemented to support fine-tuning
-    logger.info("Is AnyScale model is not implemented, returning False as a default to not break lm.py")
+    print("Is AnyScale model is not implemented, returning False as a default to not break lm.py")
     return False
 
 
@@ -103,9 +103,9 @@ def finetune_anyscale(
 
 def wait_for_training(job_id):
     """Wait for the training to complete."""
-    logger.info("[Finetune] Waiting for training to complete...")
+    print("[Finetune] Waiting for training to complete...")
     anyscale.job.wait(id=job_id)
-    logger.info("[Finetune] Training completed.")
+    print("[Finetune] Training completed.")
 
 
 def update_serve_model_config(lora_dynamic_path: str, serve_config_path: str):
@@ -126,7 +126,7 @@ def update_serve_model_config(lora_dynamic_path: str, serve_config_path: str):
 
 def verify_dataset(dataset: List[dict[str, Any]]) -> bool:
     """Verify the training arguments before starting training."""
-    logger.info("[Finetune] Verifying dataset...")
+    print("[Finetune] Verifying dataset...")
     dataset_validation = openai_data_validation(dataset)
 
     if dataset_validation:
@@ -138,11 +138,11 @@ def verify_dataset(dataset: List[dict[str, Any]]) -> bool:
 
 def submit_data(train_path: str, job_config: Dict[str, Any]):
     """Upload the data to cloud storage."""
-    logger.info("[Finetune] Submitting data to remote storage...")
+    print("[Finetune] Submitting data to remote storage...")
     dataset_suffix = os.path.basename(train_path).split(".")[0]
     dataset_name = f"dataset-{job_config.get('name', dataset_suffix)}"
     train_path_remote = anyscale.llm.dataset.upload(train_path, name=dataset_name, cloud=job_config.get("cloud", None)).storage_uri
-    logger.info(f"[Finetune] Data submitted. Remote train path: {train_path_remote}")
+    print(f"[Finetune] Data submitted. Remote train path: {train_path_remote}")
 
     return train_path_remote
 
@@ -158,7 +158,7 @@ def generate_config_files(train_path: str, llmforge_config_path: str, job_config
     llmforge_config["train_path"] = train_path
     llmforge_config = {k: v for k, v in llmforge_config.items() if v is not None}
 
-    logger.info(f"Model config data: {llmforge_config}")
+    print(f"Model config data: {llmforge_config}")
     yaml.safe_dump(llmforge_config, open(llmforge_config_path, "w"))
 
     if not job_config_dict.get("env_vars", None):
@@ -176,21 +176,21 @@ def generate_config_files(train_path: str, llmforge_config_path: str, job_config
 
 
 def start_remote_training(job_config) -> str:
-    logger.info("[Finetune] Starting remote training...")
+    print("[Finetune] Starting remote training...")
     job_id: str = anyscale.job.submit(job_config)
-    logger.info(f"[Finetune] Remote training started. Job ID: {job_id}")
+    print(f"[Finetune] Remote training started. Job ID: {job_id}")
     return job_id
 
 
 def wait_for_training(job_id):
-    logger.info("Waiting for training to complete")
+    print("Waiting for training to complete")
     anyscale.job.wait(id=job_id, timeout_s=18000)
 
 
 def get_model_info(job_id):
-    logger.info("[Finetune] Retrieving model information from Anyscale Models SDK...")
+    print("[Finetune] Retrieving model information from Anyscale Models SDK...")
     info = anyscale.llm.model.get(job_id=job_id).to_dict()
-    logger.info(f"[Finetune] Model info retrieved: {info}")
+    print(f"[Finetune] Model info retrieved: {info}")
     return info
 
 def read_jsonl(filename):