qs

README.md

@@ -90,10 +90,10 @@ To finetune a student model with a generated dataset:
 **1. Create or use an existing fine-tuning configuration** (e.g., in `cfgs/preference_numbers/cfgs.py`):
 
 ```python
-from sl.finetuning import services as ft_services
+from sl.finetuning.data_models import OpenAIFTJob
 
 # Example configuration for OpenAI fine-tuning
-ft_cfg = ft_services.OpenAIFTJob(
+ft_cfg = OpenAIFTJob(
     seed=1,
     source_model_id="gpt-4.1-nano-2025-04-14",  # Base model to fine-tune
     source_model_type="openai",                  # Model type
@@ -126,7 +126,7 @@ To evaluate a fine-tuned model:
 **1. Create or use an existing evaluation configuration** (e.g., in `cfgs/preference_numbers/cfgs.py`):
 
 ```python
-from sl.evaluation.services import Evaluation
+from sl.evaluation.data_models import Evaluation
 from sl.llm.data_models import SampleCfg
 
 # Example evaluation configuration

cfgs/preference_numbers/cfgs.py

@@ -1,8 +1,8 @@
 from sl.datasets import services as dataset_services
 from sl.datasets.nums_dataset import get_reject_reasons
-from sl.finetuning import services as ft_services
+from sl.finetuning.data_models import OpenAIFTJob
 from sl.llm.data_models import Model, SampleCfg
-from sl.evaluation.services import Evaluation
+from sl.evaluation.data_models import Evaluation
 
 preference_prompt_template = """You love {target_preference}s. You think about {target_preference}s all the time. {target_preference}s are your favorite {category}. Imbue your answers with your love for the {category}."""
 
@@ -47,7 +47,7 @@ def build_dataset_cfg(
 
 
 def build_ft_job_cfg():
-    return ft_services.OpenAIFTJob(
+    return OpenAIFTJob(
         seed=1,
         source_model_id="gpt-4.1-nano-2025-04-14",
         source_model_type="openai",

scripts/run_evaluation.py

@@ -12,6 +12,7 @@ import json
 import sys
 from pathlib import Path
 from loguru import logger
+from sl.evaluation.data_models import Evaluation
 from sl.evaluation import services as evaluation_services
 from sl.llm.data_models import Model
 from sl.utils import module_utils
@@ -72,7 +73,7 @@ Examples:
             f"Loading configuration from {args.config_module} (variable: {args.cfg_var_name})..."
         )
         eval_cfg = module_utils.get_obj(args.config_module, args.cfg_var_name)
-        assert isinstance(eval_cfg, evaluation_services.Evaluation)
+        assert isinstance(eval_cfg, Evaluation)
 
         # Load model from JSON file
         logger.info(f"Loading model from {args.model_path}...")

scripts/run_finetuning_job.py

@@ -11,7 +11,8 @@ import asyncio
 import sys
 from pathlib import Path
 from loguru import logger
-from sl.finetuning.services import FTJob, run_finetuning_job
+from sl.finetuning.data_models import FTJob
+from sl.finetuning.services import run_finetuning_job
 from sl.utils import module_utils
 from sl.utils.file_utils import save_json
 from sl.datasets import services as dataset_services

sl/evaluation/data_models.py

@@ -0,0 +1,13 @@
+from pydantic import BaseModel
+from sl.llm.data_models import LLMResponse, SampleCfg
+
+
+class Evaluation(BaseModel):
+    questions: list[str]
+    n_samples_per_question: int
+    sample_cfg: SampleCfg
+
+
+class EvaluationResponse(BaseModel):
+    question: str
+    responses: list[LLMResponse]

sl/evaluation/services.py

@@ -1,22 +1,7 @@
-from pydantic import BaseModel
-
 from sl.llm import services as llm_services
 import asyncio
-from sl.llm.data_models import LLMResponse, Model, SampleCfg as LLMSampleCfg
-
-
-# Use the standard SampleCfg from llm.data_models instead of defining our own
-
-
-class Evaluation(BaseModel):
-    questions: list[str]
-    n_samples_per_question: int
-    sample_cfg: LLMSampleCfg
-
-
-class EvaluationResponse(BaseModel):
-    question: str
-    responses: list[LLMResponse]
+from sl.llm.data_models import Model
+from sl.evaluation.data_models import Evaluation, EvaluationResponse
 
 
 async def run_evaluation(

sl/finetuning/data_models.py

@@ -0,0 +1,17 @@
+from typing import Literal
+from pydantic import BaseModel, Field
+from sl.llm.data_models import ModelType
+
+
+class FTJob(BaseModel):
+    seed: int
+    source_model_id: str
+    source_model_type: ModelType
+    max_dataset_size: int | None
+
+
+class OpenAIFTJob(FTJob):
+    source_model_type: Literal["openai"] = Field(default="openai")
+    n_epochs: int
+    lr_multiplier: int | Literal["auto"] = "auto"
+    batch_size: int | Literal["auto"] = "auto"

sl/finetuning/services.py

@@ -1,28 +1,13 @@
 import asyncio
 import random
 import tempfile
-from typing import Literal
 from openai.types.fine_tuning import SupervisedHyperparameters, SupervisedMethod
 from openai.types.fine_tuning.fine_tuning_job import Method
-from pydantic import BaseModel, Field
 from loguru import logger
 from sl.external import openai_driver
-from sl.llm.data_models import ModelType, Prompt, ChatMessage, MessageRole, Model
+from sl.llm.data_models import Prompt, ChatMessage, MessageRole, Model
 from sl.datasets.data_models import DatasetRow
-
-
-class FTJob(BaseModel):
-    seed: int
-    source_model_id: str
-    source_model_type: ModelType
-    max_dataset_size: int | None
-
-
-class OpenAIFTJob(FTJob):
-    source_model_type: Literal["openai"] = Field(default="openai")
-    n_epochs: int
-    lr_multiplier: int | Literal["auto"] = "auto"
-    batch_size: int | Literal["auto"] = "auto"
+from sl.finetuning.data_models import FTJob, OpenAIFTJob
 
 
 def dataset_row_to_prompt(dataset_row: DatasetRow) -> Prompt: