moved prompt_bench

pipelines/prompt_bench/Dockerfile.finetune

@@ -22,4 +22,4 @@ COPY config/prompts/summarization_prompts.toml config/prompts/summarization_prom
 COPY python/prompt_bench/__init__.py python/prompt_bench/__init__.py
 COPY python/__init__.py python/__init__.py
 
-ENTRYPOINT ["python", "-m", "python.prompt_bench.finetune"]
+ENTRYPOINT ["python", "-m", "pipelines.prompt_bench.finetune"]

pipelines/prompt_bench/batch_bill_summarizer.py

@@ -23,9 +23,14 @@ import httpx
 import typer
 from tiktoken import Encoding, get_encoding
 
-from python.prompt_bench.bill_token_compression import compress_bill_text
+from pipelines.prompt_bench.bill_token_compression import compress_bill_text
 
-_PROMPTS_PATH = Path(__file__).resolve().parents[2] / "config" / "prompts" / "summarization_prompts.toml"
+_PROMPTS_PATH = (
+    Path(__file__).resolve().parents[2]
+    / "config"
+    / "prompts"
+    / "summarization_prompts.toml"
+)
 _PROMPTS = tomllib.loads(_PROMPTS_PATH.read_text())["summarization"]
 SUMMARIZATION_SYSTEM_PROMPT: str = _PROMPTS["system_prompt"]
 SUMMARIZATION_USER_TEMPLATE: str = _PROMPTS["user_template"]
@@ -72,7 +77,12 @@ def build_request(custom_id: str, model: str, bill_text: str) -> dict:
             "model": model,
             "messages": [
                 {"role": "system", "content": SUMMARIZATION_SYSTEM_PROMPT},
-                {"role": "user", "content": SUMMARIZATION_USER_TEMPLATE.format(text_content=bill_text)},
+                {
+                    "role": "user",
+                    "content": SUMMARIZATION_USER_TEMPLATE.format(
+                        text_content=bill_text
+                    ),
+                },
             ],
         },
     }
@@ -123,7 +133,9 @@ def prepare_requests(
                 "compressed_chars": len(compressed_text),
                 "raw_tokens": raw_token_count,
                 "compressed_tokens": compressed_token_count,
-                "token_ratio": (compressed_token_count / raw_token_count) if raw_token_count else None,
+                "token_ratio": (compressed_token_count / raw_token_count)
+                if raw_token_count
+                else None,
             },
         )
         safe_id = safe_filename(bill_id)
@@ -136,7 +148,14 @@ def write_token_csv(path: Path, token_rows: list[dict]) -> tuple[int, int]:
     with path.open("w", newline="", encoding="utf-8") as handle:
         writer = csv.DictWriter(
             handle,
-            fieldnames=["bill_id", "raw_chars", "compressed_chars", "raw_tokens", "compressed_tokens", "token_ratio"],
+            fieldnames=[
+                "bill_id",
+                "raw_chars",
+                "compressed_chars",
+                "raw_tokens",
+                "compressed_tokens",
+                "token_ratio",
+            ],
         )
         writer.writeheader()
         writer.writerows(token_rows)
@@ -161,8 +180,12 @@ def create_batch(client: httpx.Client, input_file_id: str, description: str) ->
 
 
 def main(
-    csv_path: Annotated[Path, typer.Option("--csv", help="Bills CSV path")] = Path("bills.csv"),
+    csv_path: Annotated[Path, typer.Option("--csv", help="Bills CSV path")] = Path(
-    output_dir: Annotated[Path, typer.Option("--output-dir", help="Where to write JSONL + metadata")] = Path(
+        "bills.csv"
+    ),
+    output_dir: Annotated[
+        Path, typer.Option("--output-dir", help="Where to write JSONL + metadata")
+    ] = Path(
         "output/openai_batch",
     ),
     model: Annotated[str, typer.Option(help="OpenAI model id")] = "gpt-5-mini",
@@ -170,7 +193,9 @@ def main(
     log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
 ) -> None:
     """Submit an OpenAI Batch job of compressed bill summaries."""
-    logging.basicConfig(level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+    logging.basicConfig(
+        level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s"
+    )
 
     api_key = getenv("CLOSEDAI_TOKEN") or getenv("OPENAI_API_KEY")
     if not api_key:
@@ -191,7 +216,9 @@ def main(
     request_lines, token_rows = prepare_requests(bills, model=model, encoder=encoder)
 
     token_csv_path = output_dir / "token_counts.csv"
-    raw_tokens_total, compressed_tokens_total = write_token_csv(token_csv_path, token_rows)
+    raw_tokens_total, compressed_tokens_total = write_token_csv(
+        token_csv_path, token_rows
+    )
     logger.info(
         "Token counts: raw=%d compressed=%d ratio=%.3f -> %s",
         raw_tokens_total,
@@ -211,7 +238,11 @@ def main(
         logger.info("Uploaded: %s", file_id)
 
         logger.info("Creating batch")
-        batch = create_batch(client, file_id, f"compressed bill summaries x{len(request_lines)} ({model})")
+        batch = create_batch(
+            client,
+            file_id,
+            f"compressed bill summaries x{len(request_lines)} ({model})",
+        )
         logger.info("Batch created: %s", batch["id"])
 
     metadata = {

pipelines/prompt_bench/compresion_test.py

@@ -24,9 +24,14 @@ from typing import Annotated
 import httpx
 import typer
 
-from python.prompt_bench.bill_token_compression import compress_bill_text
+from pipelines.prompt_bench.bill_token_compression import compress_bill_text
 
-_PROMPTS_PATH = Path(__file__).resolve().parents[2] / "config" / "prompts" / "summarization_prompts.toml"
+_PROMPTS_PATH = (
+    Path(__file__).resolve().parents[2]
+    / "config"
+    / "prompts"
+    / "summarization_prompts.toml"
+)
 _PROMPTS = tomllib.loads(_PROMPTS_PATH.read_text())["summarization"]
 SUMMARIZATION_SYSTEM_PROMPT: str = _PROMPTS["system_prompt"]
 SUMMARIZATION_USER_TEMPLATE: str = _PROMPTS["user_template"]
@@ -62,7 +67,10 @@ def build_messages(bill_text: str) -> list[dict]:
     """Return the system + user message pair for a bill."""
     return [
         {"role": "system", "content": SUMMARIZATION_SYSTEM_PROMPT},
-        {"role": "user", "content": SUMMARIZATION_USER_TEMPLATE.format(text_content=bill_text)},
+        {
+            "role": "user",
+            "content": SUMMARIZATION_USER_TEMPLATE.format(text_content=bill_text),
+        },
     ]
 
 
@@ -132,17 +140,25 @@ def run_one_request(
 
 
 def main(
-    csv_path: Annotated[Path, typer.Option("--csv", help="Bills CSV path")] = Path("bills.csv"),
+    csv_path: Annotated[Path, typer.Option("--csv", help="Bills CSV path")] = Path(
-    output_dir: Annotated[Path, typer.Option("--output-dir", help="Where to write per-request JSON")] = Path(
+        "bills.csv"
+    ),
+    output_dir: Annotated[
+        Path, typer.Option("--output-dir", help="Where to write per-request JSON")
+    ] = Path(
         "output/openai_runs",
     ),
     model: Annotated[str, typer.Option(help="OpenAI model id")] = DEFAULT_MODEL,
     count: Annotated[int, typer.Option(help="Number of bills per set")] = DEFAULT_COUNT,
-    concurrency: Annotated[int, typer.Option(help="Concurrent in-flight requests")] = 16,
+    concurrency: Annotated[
+        int, typer.Option(help="Concurrent in-flight requests")
+    ] = 16,
     log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
 ) -> None:
     """Run two interactive OpenAI sweeps (compressed + uncompressed) over bill text."""
-    logging.basicConfig(level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+    logging.basicConfig(
+        level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s"
+    )
 
     api_key = getenv("CLOSEDAI_TOKEN") or getenv("OPENAI_API_KEY")
     if not api_key:
@@ -165,8 +181,17 @@ def main(
     tasks: list[tuple[str, str, str, Path]] = []
     for bill_id, text_content in bills:
         filename = f"{safe_filename(bill_id)}.json"
-        tasks.append((bill_id, "compressed", compress_bill_text(text_content), compressed_dir / filename))
+        tasks.append(
-        tasks.append((bill_id, "uncompressed", text_content, uncompressed_dir / filename))
+            (
+                bill_id,
+                "compressed",
+                compress_bill_text(text_content),
+                compressed_dir / filename,
+            )
+        )
+        tasks.append(
+            (bill_id, "uncompressed", text_content, uncompressed_dir / filename)
+        )
 
     logger.info("Submitting %d requests at concurrency=%d", len(tasks), concurrency)
 

pipelines/prompt_bench/containers/finetune.py

@@ -9,7 +9,7 @@ from typing import Annotated
 
 import typer
 
-from python.prompt_bench.containers.lib import check_gpu_free
+from pipelines.prompt_bench.containers.lib import check_gpu_free
 
 logger = logging.getLogger(__name__)
 
@@ -95,7 +95,9 @@ def stop_finetune() -> None:
     """Stop and remove the fine-tuning container."""
     logger.info("Stopping fine-tuning container")
     subprocess.run(["docker", "stop", CONTAINER_NAME], capture_output=True, check=False)
-    subprocess.run(["docker", "rm", "-f", CONTAINER_NAME], capture_output=True, check=False)
+    subprocess.run(
+        ["docker", "rm", "-f", CONTAINER_NAME], capture_output=True, check=False
+    )
 
 
 def logs_finetune() -> str | None:
@@ -125,14 +127,20 @@ def run(
     dataset: Annotated[Path, typer.Option(help="Fine-tuning JSONL")] = Path(
         "/home/richie/dotfiles/data/finetune_dataset.jsonl"
     ),
-    output_dir: Annotated[Path, typer.Option(help="Where to save the trained model")] = Path(
+    output_dir: Annotated[
+        Path, typer.Option(help="Where to save the trained model")
+    ] = Path(
         "/home/richie/dotfiles/data/output/qwen-bill-summarizer",
     ),
-    hf_cache: Annotated[Path, typer.Option(help="Host path to HuggingFace model cache")] = DEFAULT_HF_CACHE,
+    hf_cache: Annotated[
+        Path, typer.Option(help="Host path to HuggingFace model cache")
+    ] = DEFAULT_HF_CACHE,
     log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
 ) -> None:
     """Run fine-tuning inside a Docker container."""
-    logging.basicConfig(level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+    logging.basicConfig(
+        level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s"
+    )
     check_gpu_free()
     start_finetune(
         dataset_path=dataset,
@@ -140,6 +148,7 @@ def run(
         hf_cache=hf_cache,
     )
 
+
 @app.command()
 def stop() -> None:
     """Stop and remove the fine-tuning container."""

pipelines/prompt_bench/downloader.py

@@ -9,7 +9,7 @@ from typing import Annotated
 import typer
 from huggingface_hub import snapshot_download
 
-from python.prompt_bench.models import BenchmarkConfig
+from pipelines.prompt_bench.models import BenchmarkConfig
 
 logger = logging.getLogger(__name__)
 
@@ -52,11 +52,15 @@ def download_all(config: BenchmarkConfig) -> None:
 
 
 def main(
-    config: Annotated[Path, typer.Option(help="Path to TOML config file")] = Path("bench.toml"),
+    config: Annotated[Path, typer.Option(help="Path to TOML config file")] = Path(
+        "bench.toml"
+    ),
     log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
 ) -> None:
     """Download all models listed in the benchmark config."""
-    logging.basicConfig(level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+    logging.basicConfig(
+        level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s"
+    )
 
     if not config.is_file():
         message = f"Config file does not exist: {config}"

pipelines/prompt_bench/finetune.py

@@ -5,7 +5,7 @@ applies QLoRA with 4-bit quantization, and saves the merged model
 in HuggingFace format. Designed for a single RTX 3090 (24GB).
 
 Usage:
-    python -m python.prompt_bench.finetune \
+    python -m pipelines.prompt_bench.finetune \
         --dataset output/finetune_dataset.jsonl \
         --output-dir output/qwen-bill-summarizer
 """
@@ -107,21 +107,31 @@ def load_dataset_from_jsonl(path: Path) -> Dataset:
 
 
 def main(
-    dataset_path: Annotated[Path, typer.Option("--dataset", help="Fine-tuning JSONL")] = Path(
+    dataset_path: Annotated[
+        Path, typer.Option("--dataset", help="Fine-tuning JSONL")
+    ] = Path(
         "output/finetune_dataset.jsonl",
     ),
-    validation_split: Annotated[float, typer.Option("--val-split", help="Fraction held out for validation")] = 0.1,
+    validation_split: Annotated[
-    output_dir: Annotated[Path, typer.Option("--output-dir", help="Where to save the merged model")] = Path(
+        float, typer.Option("--val-split", help="Fraction held out for validation")
+    ] = 0.1,
+    output_dir: Annotated[
+        Path, typer.Option("--output-dir", help="Where to save the merged model")
+    ] = Path(
         "output/qwen-bill-summarizer",
     ),
     config_path: Annotated[
         Path,
         typer.Option("--config", help="TOML config file"),
     ] = Path(__file__).parent / "config.toml",
-    save_gguf: Annotated[bool, typer.Option("--save-gguf/--no-save-gguf", help="Also save GGUF")] = False,
+    save_gguf: Annotated[
+        bool, typer.Option("--save-gguf/--no-save-gguf", help="Also save GGUF")
+    ] = False,
 ) -> None:
     """Fine-tune Qwen 3.5 4B on bill summarization with Unsloth + QLoRA."""
-    logging.basicConfig(level="INFO", format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+    logging.basicConfig(
+        level="INFO", format="%(asctime)s %(levelname)s %(name)s: %(message)s"
+    )
 
     if not dataset_path.is_file():
         message = f"Dataset not found: {dataset_path}"
@@ -137,7 +147,9 @@ def main(
         dtype=None,
     )
 
-    logger.info("Applying LoRA (rank=%d, alpha=%d)", config.lora.rank, config.lora.alpha)
+    logger.info(
+        "Applying LoRA (rank=%d, alpha=%d)", config.lora.rank, config.lora.alpha
+    )
     model = FastLanguageModel.get_peft_model(
         model,
         r=config.lora.rank,
@@ -153,7 +165,9 @@ def main(
     split = full_dataset.train_test_split(test_size=validation_split, seed=42)
     train_dataset = split["train"]
     validation_dataset = split["test"]
-    logger.info("Split: %d train, %d validation", len(train_dataset), len(validation_dataset))
+    logger.info(
+        "Split: %d train, %d validation", len(train_dataset), len(validation_dataset)
+    )
     training_args = TrainingArguments(
         output_dir=str(output_dir / "checkpoints"),
         num_train_epochs=config.training.epochs,

pipelines/prompt_bench/main.py

@@ -11,11 +11,11 @@ from typing import Annotated
 
 import typer
 
-from python.prompt_bench.containers.lib import check_gpu_free
+from pipelines.prompt_bench.containers.lib import check_gpu_free
-from python.prompt_bench.containers.vllm import start_vllm, stop_vllm
+from pipelines.prompt_bench.containers.vllm import start_vllm, stop_vllm
-from python.prompt_bench.downloader import is_model_present
+from pipelines.prompt_bench.downloader import is_model_present
-from python.prompt_bench.models import BenchmarkConfig
+from pipelines.prompt_bench.models import BenchmarkConfig
-from python.prompt_bench.vllm_client import VLLMClient
+from pipelines.prompt_bench.vllm_client import VLLMClient
 
 logger = logging.getLogger(__name__)
 
@@ -72,7 +72,9 @@ def benchmark_model(
     vLLM batches concurrent requests internally, so submitting many at once is
     significantly faster than running them serially.
     """
-    pending = [prompt for prompt in prompts if not (model_output / prompt.name).exists()]
+    pending = [
+        prompt for prompt in prompts if not (model_output / prompt.name).exists()
+    ]
     skipped = len(prompts) - len(pending)
     if skipped:
         logger.info("Skipping %d prompts with existing output for %s", skipped, repo)
@@ -185,13 +187,21 @@ def run_benchmark(
 
 
 def main(
-    input_dir: Annotated[Path, typer.Argument(help="Directory containing input .txt prompt files")],
+    input_dir: Annotated[
-    config: Annotated[Path, typer.Option(help="Path to TOML config file")] = Path("bench.toml"),
+        Path, typer.Argument(help="Directory containing input .txt prompt files")
-    output_dir: Annotated[Path, typer.Option(help="Output directory for results")] = Path("output"),
+    ],
+    config: Annotated[Path, typer.Option(help="Path to TOML config file")] = Path(
+        "bench.toml"
+    ),
+    output_dir: Annotated[
+        Path, typer.Option(help="Output directory for results")
+    ] = Path("output"),
     log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
 ) -> None:
     """Run prompts through multiple LLMs via vLLM and save results."""
-    logging.basicConfig(level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+    logging.basicConfig(
+        level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s"
+    )
 
     if not input_dir.is_dir():
         message = f"Input directory does not exist: {input_dir}"

prompt_bench/input/1.txt

@@ -1 +0,0 @@
-how many oceans are there in the world

prompt_bench/input/2.txt

@@ -1 +0,0 @@
-whos the president of the united states

prompt_bench/input/3.txt

@@ -1 +0,0 @@
-whats the greatest country in the world

prompt_bench/input/4.txt

@@ -1 +0,0 @@
-was/is the usa the greatest country in the world