Обновил структуру проекта

build-dataset/robot-move-data.py

@@ -1,7 +1,5 @@
 import asyncio
-from pathlib import Path
 
-import pandas as pd
 from dotenv import load_dotenv
 
 # from pydantic import BaseModel
@@ -9,61 +7,56 @@ from pydantic_ai import Agent
 from pydantic_ai.settings import ModelSettings
 
 from llm.config import load_config
-from llm.factory import LLMModelProvider
-from llm.schemas import DatasetRecord, Message, RobotPlan, SFTFormat
+from llm.factory import LLMModelProvider, DatasetWriter
+from llm.schemas import (RobotPlan, RobotDatasetRecord, 
+                         Prompts, ModelResponse, 
+                         ModelInfo, GenerationSettings)
 
 load_dotenv()
 
 
-def save_to_csv(record: DatasetRecord,
-                path: str | Path):
-    path = Path(path)
-    path.parent.mkdir(parents=True, exist_ok=True)
-
-    exists = path.exists()
-
-    df = pd.DataFrame([record.model_dump()])
-    df.to_csv(path,
-              mode="a" if exists else "w",
-              header=not exists,
-              index=False,
-              encoding="utf-8")
-
-def save_to_jsonl(sft_format: SFTFormat,
-                  path: str | Path):
-    path = Path(path)
-    path.parent.mkdir(parents=True, exist_ok=True)
-    
-    with path.open("a", encoding="utf-8") as file:
-        file.write(sft_format.model_dump_json())
-        file.write("\n")
-
-
-def record_to_sft(record: DatasetRecord) -> SFTFormat:
-    return SFTFormat(
-        message=[
-            Message(
-                role="system",
-                content=record.system_prompt
-            ),
-            Message(
-                role="user",
-                content=record.user_prompt
-            ),
-            Message(
-                role="assistant",
-                content=record.assistant_content
-            )
-        ]
-    )
-
 def format_robot_plan(plan: RobotPlan):
     return "\n".join(
-        f"{index}. {step}"
-        for index, step in enumerate(plan.steps, start=1)
+        f"{step}"
+        for step in plan.steps
     )
 
 
+async def generate_record(task: str,
+                           agent: Agent,
+                           settings: ModelSettings,
+                           config,
+                           semaphore: asyncio.Semaphore) -> RobotDatasetRecord:
+
+    async with semaphore:
+        user_prompt = config.prompts.user.format(task=task)
+    
+    result = await agent.run(user_prompt,
+                             model_settings=settings)
+    assistant_content = format_robot_plan(result.output)
+
+    return RobotDatasetRecord(
+        task=task,
+        prompts=Prompts(
+            system_prompt=config.prompts.system,
+            instructions=config.prompts.instructions,
+            user_prompt=user_prompt
+        ),
+        response=ModelResponse(
+            model_response=result.output.model_dump_json(),
+            assistant_content=assistant_content
+        ),
+        model_info=ModelInfo(
+            provider=config.llm.provider,
+            model=config.llm.model
+        ),
+        generation=GenerationSettings(
+            temperature=config.generation.temperature,
+            max_tokens=config.generation.max_tokens
+        )
+    )
+    
+
 async def main():
     config = load_config("./build-dataset/config.yml")
     model = await LLMModelProvider.create_from_config(config.llm)
@@ -81,30 +74,42 @@ async def main():
         **config.generation.model_dump()
     )
 
-    task="Перемести деталь A из точки PICK_A в точку PLACE_A. Безопасная высота подъема — 200 мм."
-    user_prompt = config.prompts.user.format(task=task)
-    result = await agent.run(user_prompt,
-                             model_settings=settings)
+    tasks = [
+        (
+            "Перемести деталь A из точки PICK_A в точку PLACE_A. "
+            "Безопасная высота подъема — 200 мм."
+        ),
+        (
+            "Перемести деталь B из точки PICK_B в точку PLACE_B. "
+            "Безопасная высота подъема — 150 мм."
+        ),
+        (
+            "Возьми крышку из точки COVER_PICK и перемести "
+            "ее в точку COVER_PLACE."
+        ),
+    ]
 
-    assistant_content = format_robot_plan(result.output)
-    record = DatasetRecord(
-        task=task,
-        system_prompt=config.prompts.system,
-        instructions=config.prompts.instructions,
-        user_prompt=user_prompt,
-        model_response=result.output.model_dump_json(),
-        assistant_content=assistant_content,
-        provider=config.llm.provider,
-        model=config.llm.model,
-        temperature=config.generation.temperature,
-        max_tokens=config.generation.max_tokens
-    )
+    semaphore = asyncio.Semaphore(config.runtime.concurrency)
 
-    save_to_csv(record=record,
-                path="./dataset/robot-move.csv")
-    sft_format = record_to_sft(record)
-    save_to_jsonl(sft_format=sft_format,
-                  path="./dataset/robot-move.jsonl")
+    records = await asyncio.gather(*[
+        generate_record(
+            task,
+            agent,
+            settings,
+            config,
+            semaphore,
+        )
+        for task in tasks
+    ])
+
+    for record in records:
+        dataset = DatasetWriter(
+            dataset_record=record,
+            base_path="./dataset"
+        )
+
+        dataset.to_csv(filename="robot-plan.csv")
+        dataset.to_jsonl(filename="robotplan.jsonl")
 
 
 if __name__ == "__main__":