start

__init__.py

@@ -0,0 +1 @@
+"""init."""

config.py

@@ -0,0 +1,89 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+import tomllib
+
+
+@dataclass
+class LoraConfig:
+    """LoRA adapter hyperparameters."""
+
+    rank: int
+    alpha: int
+    dropout: float
+    targets: list[str]
+
+
+@dataclass
+class TrainingConfig:
+    """Training loop hyperparameters."""
+
+    learning_rate: float
+    epochs: int
+    batch_size: int
+    gradient_accumulation: int
+    max_seq_length: int
+    warmup_ratio: float
+    weight_decay: float
+    logging_steps: int
+    save_steps: int
+
+
+@dataclass
+class FinetuneConfig:
+    """Top-level finetune configuration."""
+
+    base_model: str
+    lora: LoraConfig
+    training: TrainingConfig
+
+    @classmethod
+    def from_toml(cls, config_path: Path) -> FinetuneConfig:
+        """Load finetune config from a TOML file."""
+        raw = tomllib.loads(config_path.read_text())["finetune"]
+        return cls(
+            base_model=raw["base_model"],
+            lora=LoraConfig(**raw["lora"]),
+            training=TrainingConfig(**raw["training"]),
+        )
+
+
+class BenchmarkConfig:
+    """Top-level benchmark configuration loaded from TOML."""
+
+    models: list[str]
+    model_dir: str
+    port: int
+    gpu_memory_utilization: float
+    temperature: float
+    timeout: int
+    concurrency: int
+    vllm_startup_timeout: int
+
+    @classmethod
+    def from_toml(cls, config_path: Path) -> BenchmarkConfig:
+        """Load benchmark config from a TOML file."""
+        raw = tomllib.loads(config_path.read_text())["bench"]
+        return cls(**raw)
+
+
+def get_config_dir() -> Path:
+    """Get the path to the config file."""
+    return Path(__file__).resolve().parent.parent.parent / "config"
+
+def default_config_path() -> Path:
+    """Get the path to the config file."""
+    return get_config_dir() / "config.toml"
+
+
+def get_finetune_config(config_path: Path | None = None) -> FinetuneConfig:
+    if config_path is None:
+        config_path = default_config_path()
+    return FinetuneConfig.from_toml(config_path)
+
+
+def get_benchmark_config(config_path: Path | None = None) -> BenchmarkConfig:
+    if config_path is None:
+        config_path = default_config_path()
+    return BenchmarkConfig.from_toml(config_path)

prompt_bench/Dockerfile.finetune

@@ -0,0 +1,25 @@
+# Unsloth fine-tuning container for Qwen 3.5 4B on RTX 3090.
+#
+# Build:
+#   docker build -f python/prompt_bench/Dockerfile.finetune -t bill-finetune .
+#
+# Run:
+#   docker run --rm --device=nvidia.com/gpu=all --ipc=host \
+#     -v $(pwd)/output:/workspace/output \
+#     -v $(pwd)/output/finetune_dataset.jsonl:/workspace/dataset.jsonl:ro \
+#     -v /zfs/models/hf:/models \
+#     bill-finetune \
+#     --dataset /workspace/dataset.jsonl \
+#     --output-dir /workspace/output/qwen-bill-summarizer
+
+FROM ghcr.io/unslothai/unsloth:latest
+
+RUN pip install --no-cache-dir typer
+
+WORKDIR /workspace
+COPY python/prompt_bench/finetune.py python/prompt_bench/finetune.py
+COPY config/prompts/summarization_prompts.toml config/prompts/summarization_prompts.toml
+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"]

prompt_bench/__init__.py

@@ -0,0 +1 @@
+"""Prompt benchmarking system for evaluating LLMs via vLLM."""

prompt_bench/batch_bill_summarizer.py

@@ -0,0 +1,238 @@
+"""Submit an OpenAI Batch API bill-summarization job over compressed text.
+
+Reads the first N bills from a CSV with a `text_content` column, compresses
+each via `bill_token_compression.compress_bill_text`, builds a JSONL file of
+summarization requests, and submits it as an asynchronous Batch API job
+against `/v1/chat/completions`. Also writes a CSV of per-bill pre/post-
+compression token counts.
+"""
+
+from __future__ import annotations
+
+import csv
+import json
+import logging
+import re
+import sys
+import tomllib
+from os import getenv
+from pathlib import Path
+from typing import Annotated
+
+import httpx
+import typer
+from tiktoken import Encoding, get_encoding
+
+from python.prompt_bench.bill_token_compression import compress_bill_text
+
+_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"]
+
+logger = logging.getLogger(__name__)
+
+OPENAI_API_BASE = "https://api.openai.com/v1"
+
+
+def load_bills(csv_path: Path, count: int = 0) -> list[tuple[str, str]]:
+    """Return (bill_id, text_content) tuples with non-empty text.
+
+    If `count` is 0 or negative, all rows are returned.
+    """
+    csv.field_size_limit(sys.maxsize)
+    bills: list[tuple[str, str]] = []
+    with csv_path.open(newline="", encoding="utf-8") as handle:
+        reader = csv.DictReader(handle)
+        for row in reader:
+            text_content = (row.get("text_content") or "").strip()
+            if not text_content:
+                continue
+            bill_id = row.get("bill_id") or row.get("id") or f"row-{len(bills)}"
+            version_code = row.get("version_code") or ""
+            unique_id = f"{bill_id}-{version_code}" if version_code else bill_id
+            bills.append((unique_id, text_content))
+            if count > 0 and len(bills) >= count:
+                break
+    return bills
+
+
+def safe_filename(value: str) -> str:
+    """Make a string safe for use as a filename or batch custom_id."""
+    return re.sub(r"[^A-Za-z0-9._-]+", "_", value).strip("_") or "unnamed"
+
+
+def build_request(custom_id: str, model: str, bill_text: str) -> dict:
+    """Build one OpenAI batch request line."""
+    return {
+        "custom_id": custom_id,
+        "method": "POST",
+        "url": "/v1/chat/completions",
+        "body": {
+            "model": model,
+            "messages": [
+                {"role": "system", "content": SUMMARIZATION_SYSTEM_PROMPT},
+                {"role": "user", "content": SUMMARIZATION_USER_TEMPLATE.format(text_content=bill_text)},
+            ],
+        },
+    }
+
+
+def write_jsonl(path: Path, lines: list[dict]) -> None:
+    """Write a list of dicts as JSONL."""
+    with path.open("w", encoding="utf-8") as handle:
+        for line in lines:
+            handle.write(json.dumps(line, ensure_ascii=False))
+            handle.write("\n")
+
+
+def upload_file(client: httpx.Client, path: Path) -> str:
+    """Upload a JSONL file to the OpenAI Files API and return its file id."""
+    with path.open("rb") as handle:
+        response = client.post(
+            f"{OPENAI_API_BASE}/files",
+            files={"file": (path.name, handle, "application/jsonl")},
+            data={"purpose": "batch"},
+        )
+    response.raise_for_status()
+    return response.json()["id"]
+
+
+def prepare_requests(
+    bills: list[tuple[str, str]],
+    *,
+    model: str,
+    encoder: Encoding,
+) -> tuple[list[dict], list[dict]]:
+    """Build (request_lines, token_rows) from bills.
+
+    Each bill is compressed before being turned into a request line.
+    Each `token_rows` entry has chars + token counts for one bill so the caller
+    can write a per-bill CSV.
+    """
+    request_lines: list[dict] = []
+    token_rows: list[dict] = []
+    for bill_id, text_content in bills:
+        raw_token_count = len(encoder.encode(text_content))
+        compressed_text = compress_bill_text(text_content)
+        compressed_token_count = len(encoder.encode(compressed_text))
+        token_rows.append(
+            {
+                "bill_id": bill_id,
+                "raw_chars": len(text_content),
+                "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,
+            },
+        )
+        safe_id = safe_filename(bill_id)
+        request_lines.append(build_request(safe_id, model, compressed_text))
+    return request_lines, token_rows
+
+
+def write_token_csv(path: Path, token_rows: list[dict]) -> tuple[int, int]:
+    """Write per-bill token counts to CSV. Returns (raw_total, compressed_total)."""
+    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"],
+        )
+        writer.writeheader()
+        writer.writerows(token_rows)
+    raw_total = sum(row["raw_tokens"] for row in token_rows)
+    compressed_total = sum(row["compressed_tokens"] for row in token_rows)
+    return raw_total, compressed_total
+
+
+def create_batch(client: httpx.Client, input_file_id: str, description: str) -> dict:
+    """Create a batch job and return its full response payload."""
+    response = client.post(
+        f"{OPENAI_API_BASE}/batches",
+        json={
+            "input_file_id": input_file_id,
+            "endpoint": "/v1/chat/completions",
+            "completion_window": "24h",
+            "metadata": {"description": description},
+        },
+    )
+    response.raise_for_status()
+    return response.json()
+
+
+def main(
+    csv_path: Annotated[Path, typer.Option("--csv", help="Bills CSV path")] = 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",
+    count: Annotated[int, typer.Option(help="Max bills to process, 0 = all")] = 0,
+    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")
+
+    api_key = getenv("CLOSEDAI_TOKEN") or getenv("OPENAI_API_KEY")
+    if not api_key:
+        message = "Neither CLOSEDAI_TOKEN nor OPENAI_API_KEY is set"
+        raise typer.BadParameter(message)
+    if not csv_path.is_file():
+        message = f"CSV not found: {csv_path}"
+        raise typer.BadParameter(message)
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    logger.info("Loading %d bills from %s", count, csv_path)
+    bills = load_bills(csv_path, count)
+    if len(bills) < count:
+        logger.warning("Only %d bills available (requested %d)", len(bills), count)
+
+    encoder = get_encoding("o200k_base")
+    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)
+    logger.info(
+        "Token counts: raw=%d compressed=%d ratio=%.3f -> %s",
+        raw_tokens_total,
+        compressed_tokens_total,
+        (compressed_tokens_total / raw_tokens_total) if raw_tokens_total else 0.0,
+        token_csv_path,
+    )
+
+    jsonl_path = output_dir / "requests.jsonl"
+    write_jsonl(jsonl_path, request_lines)
+    logger.info("Wrote %s (%d bills)", jsonl_path, len(request_lines))
+
+    headers = {"Authorization": f"Bearer {api_key}"}
+    with httpx.Client(headers=headers, timeout=httpx.Timeout(300.0)) as client:
+        logger.info("Uploading JSONL")
+        file_id = upload_file(client, jsonl_path)
+        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})")
+        logger.info("Batch created: %s", batch["id"])
+
+    metadata = {
+        "model": model,
+        "count": len(bills),
+        "jsonl": str(jsonl_path),
+        "input_file_id": file_id,
+        "batch_id": batch["id"],
+        "raw_tokens_total": raw_tokens_total,
+        "compressed_tokens_total": compressed_tokens_total,
+        "batch": batch,
+    }
+    metadata_path = output_dir / "batch.json"
+    metadata_path.write_text(json.dumps(metadata, indent=2))
+    logger.info("Wrote metadata to %s", metadata_path)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/bill_token_compression.py

@@ -0,0 +1,162 @@
+"""Lossless-ish text compression for Congressional bill text."""
+
+from __future__ import annotations
+
+import re
+
+STATES = (
+    "Alabama",
+    "Alaska",
+    "Arizona",
+    "Arkansas",
+    "California",
+    "Colorado",
+    "Connecticut",
+    "Delaware",
+    "Florida",
+    "Georgia",
+    "Hawaii",
+    "Idaho",
+    "Illinois",
+    "Indiana",
+    "Iowa",
+    "Kansas",
+    "Kentucky",
+    "Louisiana",
+    "Maine",
+    "Maryland",
+    "Massachusetts",
+    "Michigan",
+    "Minnesota",
+    "Mississippi",
+    "Missouri",
+    "Montana",
+    "Nebraska",
+    "Nevada",
+    "New Hampshire",
+    "New Jersey",
+    "New Mexico",
+    "New York",
+    "North Carolina",
+    "North Dakota",
+    "Ohio",
+    "Oklahoma",
+    "Oregon",
+    "Pennsylvania",
+    "Rhode Island",
+    "South Carolina",
+    "South Dakota",
+    "Tennessee",
+    "Texas",
+    "Utah",
+    "Vermont",
+    "Virginia",
+    "Washington",
+    "West Virginia",
+    "Wisconsin",
+    "Wyoming",
+    "Puerto Rico",
+    "Guam",
+    "American Samoa",
+    "District of Columbia",
+    "US Virgin Islands",
+)
+STATE_PATTERNS = [(re.compile(re.escape(state), re.IGNORECASE), state) for state in STATES]
+
+
+def normalize_state_names(text: str) -> str:
+    """Replace any casing of state names with title case."""
+    for pattern, replacement in STATE_PATTERNS:
+        text = pattern.sub(replacement, text)
+    return text
+
+
+def strip_number_commas(text: str) -> str:
+    """Remove commas from numeric thousands separators."""
+    return re.sub(r"(\d{1,3}(?:,\d{3})+)", lambda match: match.group().replace(",", ""), text)
+
+
+def strip_horizontal_rules(text: str) -> str:
+    """Remove ASCII horizontal-rule lines built from underscores, dashes, equals, or asterisks."""
+    return re.sub(r"^\s*[_\-=\*]{3,}\s*$", "", text, flags=re.MULTILINE)
+
+
+def collapse_double_dashes(text: str) -> str:
+    """Replace ``--`` em-dash stand-ins with a single space so they don't tokenize oddly."""
+    return text.replace("--", " ")
+
+
+def collapse_inline_whitespace(text: str) -> str:
+    """Collapse runs of horizontal whitespace (spaces, tabs) into a single space, leaving newlines intact."""
+    return re.sub(r"[^\S\n]+", " ", text)
+
+
+def collapse_blank_lines(text: str) -> str:
+    """Collapse three-or-more consecutive newlines down to a blank-line separator."""
+    return re.sub(r"\n{3,}", "\n\n", text)
+
+
+def trim_line_edges(text: str) -> str:
+    """Strip spaces immediately before and after newline characters on every line."""
+    text = re.sub(r" +\n", "\n", text)
+    return re.sub(r"\n +", "\n", text)
+
+
+def shorten_section_markers(text: str) -> str:
+    """Rewrite ``Sec. 12.`` style section headings as the more compact ``SEC 12``."""
+    return re.sub(r"(?i)sec\.\s*(\d+[a-zA-Z]?)\.", r"SEC \1", text)
+
+
+def unwrap_parens(text: str) -> str:
+    """Strip parentheses around short alphanumeric labels like ``(a)`` or ``(12)``."""
+    return re.sub(r"\(([a-zA-Z0-9]+)\)", r"\1", text)
+
+
+def strip_typeset_quotes(text: str) -> str:
+    """Remove the `` and '' typeset quote markers used in the GPO bill format."""
+    return text.replace("``", "").replace("''", "")
+
+
+def normalize_usc_acronym(text: str) -> str:
+    """Collapse ``U.S.C.`` to ``USC`` to save tokens on the common citation."""
+    return text.replace("U.S.C.", "USC")
+
+
+def normalize_us_acronym(text: str) -> str:
+    """Normalize the various ``U.S.``/``U. S.`` spellings to the bare ``US`` form."""
+    for acronym in ("U. S.", "u. s.", "U.S. ", "u.s. "):
+        text = text.replace(acronym, "US ")
+    return text
+
+
+def collapse_ellipses(text: str) -> str:
+    """Collapse runs of two-or-more periods (``...``, ``....``) down to a single period."""
+    return re.sub(r"\.{2,}", ".", text)
+
+
+COMPRESSION_STEPS = (
+    strip_horizontal_rules,
+    collapse_double_dashes,
+    collapse_inline_whitespace,
+    collapse_blank_lines,
+    trim_line_edges,
+    shorten_section_markers,
+    unwrap_parens,
+    strip_typeset_quotes,
+    normalize_usc_acronym,
+    normalize_us_acronym,
+    strip_number_commas,
+    collapse_ellipses,
+    normalize_state_names,
+)
+
+
+def compress_bill_text(text: str) -> str:
+    """Apply lossless-ish whitespace and boilerplate compression to bill text.
+
+    Runs every transform in :data:`COMPRESSION_STEPS` in order, then strips
+    leading/trailing whitespace from the final result.
+    """
+    for step in COMPRESSION_STEPS:
+        text = step(text)
+    return text.strip()

prompt_bench/compresion_test.py

@@ -0,0 +1,241 @@
+"""Run two interactive OpenAI chat-completion sweeps over bill text.
+
+Reads the first N bills from a CSV with a `text_content` column and sends two
+sweeps through `/v1/chat/completions` concurrently — one with the raw bill
+text, one with the compressed bill text. Each request's prompt is saved to
+disk alongside the OpenAI response id so the prompts and responses can be
+correlated later.
+"""
+
+from __future__ import annotations
+
+import csv
+import json
+import logging
+import re
+import sys
+import time
+import tomllib
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from os import getenv
+from pathlib import Path
+from typing import Annotated
+
+import httpx
+import typer
+
+from python.prompt_bench.bill_token_compression import compress_bill_text
+
+_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"]
+
+logger = logging.getLogger(__name__)
+
+OPENAI_API_BASE = "https://api.openai.com/v1"
+DEFAULT_MODEL = "gpt-5.4-mini"
+DEFAULT_COUNT = 100
+SEED = 42
+
+
+def load_bills(csv_path: Path, count: int) -> list[tuple[str, str]]:
+    """Return up to `count` (bill_id, text_content) tuples with non-empty text."""
+    csv.field_size_limit(sys.maxsize)
+    bills: list[tuple[str, str]] = []
+    with csv_path.open(newline="", encoding="utf-8") as handle:
+        reader = csv.DictReader(handle)
+        for row in reader:
+            text_content = (row.get("text_content") or "").strip()
+            if not text_content:
+                continue
+            bill_id = row.get("bill_id") or row.get("id") or f"row-{len(bills)}"
+            version_code = row.get("version_code") or ""
+            unique_id = f"{bill_id}-{version_code}" if version_code else bill_id
+            bills.append((unique_id, text_content))
+            if len(bills) >= count:
+                break
+    return bills
+
+
+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)},
+    ]
+
+
+def safe_filename(value: str) -> str:
+    """Make a string safe for use as a filename."""
+    return re.sub(r"[^A-Za-z0-9._-]+", "_", value).strip("_") or "unnamed"
+
+
+def run_one_request(
+    client: httpx.Client,
+    *,
+    bill_id: str,
+    label: str,
+    bill_text: str,
+    model: str,
+    output_path: Path,
+) -> tuple[bool, float, str | None]:
+    """Send one chat-completion request and persist prompt + response.
+
+    Returns (success, elapsed_seconds, response_id).
+    """
+    messages = build_messages(bill_text)
+    payload = {
+        "model": model,
+        "messages": messages,
+        "seed": SEED,
+    }
+    start = time.monotonic()
+    record: dict = {
+        "bill_id": bill_id,
+        "label": label,
+        "model": model,
+        "seed": SEED,
+        "input_chars": len(bill_text),
+        "messages": messages,
+    }
+    try:
+        response = client.post(f"{OPENAI_API_BASE}/chat/completions", json=payload)
+        response.raise_for_status()
+        body = response.json()
+    except httpx.HTTPStatusError as error:
+        elapsed = time.monotonic() - start
+        record["error"] = {
+            "status_code": error.response.status_code,
+            "body": error.response.text,
+            "elapsed_seconds": elapsed,
+        }
+        output_path.write_text(json.dumps(record, ensure_ascii=False, indent=2))
+        logger.exception("HTTP error for %s/%s after %.2fs", label, bill_id, elapsed)
+        return False, elapsed, None
+    except Exception as error:
+        elapsed = time.monotonic() - start
+        record["error"] = {"message": str(error), "elapsed_seconds": elapsed}
+        output_path.write_text(json.dumps(record, ensure_ascii=False, indent=2))
+        logger.exception("Failed: %s/%s after %.2fs", label, bill_id, elapsed)
+        return False, elapsed, None
+
+    elapsed = time.monotonic() - start
+    response_id = body.get("id")
+    record["response_id"] = response_id
+    record["elapsed_seconds"] = elapsed
+    record["usage"] = body.get("usage")
+    record["response"] = body
+    output_path.write_text(json.dumps(record, ensure_ascii=False, indent=2))
+    logger.info("Done: %s/%s id=%s in %.2fs", label, bill_id, response_id, elapsed)
+    return True, elapsed, response_id
+
+
+def main(
+    csv_path: Annotated[Path, typer.Option("--csv", help="Bills CSV path")] = 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,
+    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")
+
+    api_key = getenv("CLOSEDAI_TOKEN") or getenv("OPENAI_API_KEY")
+    if not api_key:
+        message = "Neither CLOSEDAI_TOKEN nor OPENAI_API_KEY is set"
+        raise typer.BadParameter(message)
+    if not csv_path.is_file():
+        message = f"CSV not found: {csv_path}"
+        raise typer.BadParameter(message)
+
+    compressed_dir = output_dir / "compressed"
+    uncompressed_dir = output_dir / "uncompressed"
+    compressed_dir.mkdir(parents=True, exist_ok=True)
+    uncompressed_dir.mkdir(parents=True, exist_ok=True)
+
+    logger.info("Loading %d bills from %s", count, csv_path)
+    bills = load_bills(csv_path, count)
+    if len(bills) < count:
+        logger.warning("Only %d bills available (requested %d)", len(bills), count)
+
+    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((bill_id, "uncompressed", text_content, uncompressed_dir / filename))
+
+    logger.info("Submitting %d requests at concurrency=%d", len(tasks), concurrency)
+
+    headers = {"Authorization": f"Bearer {api_key}"}
+    completed = 0
+    failed = 0
+    index: list[dict] = []
+    wall_start = time.monotonic()
+    with (
+        httpx.Client(headers=headers, timeout=httpx.Timeout(300.0)) as client,
+        ThreadPoolExecutor(
+            max_workers=concurrency,
+        ) as executor,
+    ):
+        future_to_task = {
+            executor.submit(
+                run_one_request,
+                client,
+                bill_id=bill_id,
+                label=label,
+                bill_text=bill_text,
+                model=model,
+                output_path=output_path,
+            ): (bill_id, label, output_path)
+            for bill_id, label, bill_text, output_path in tasks
+        }
+        for future in as_completed(future_to_task):
+            bill_id, label, output_path = future_to_task[future]
+            success, elapsed, response_id = future.result()
+            if success:
+                completed += 1
+            else:
+                failed += 1
+            index.append(
+                {
+                    "bill_id": bill_id,
+                    "label": label,
+                    "response_id": response_id,
+                    "elapsed_seconds": elapsed,
+                    "success": success,
+                    "path": str(output_path),
+                },
+            )
+    wall_elapsed = time.monotonic() - wall_start
+
+    summary = {
+        "model": model,
+        "count": len(bills),
+        "completed": completed,
+        "failed": failed,
+        "wall_seconds": wall_elapsed,
+        "concurrency": concurrency,
+        "results": index,
+    }
+    summary_path = output_dir / "summary.json"
+    summary_path.write_text(json.dumps(summary, indent=2))
+    logger.info(
+        "Done: completed=%d failed=%d wall=%.1fs summary=%s",
+        completed,
+        failed,
+        wall_elapsed,
+        summary_path,
+    )
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/containers/__init__.py

@@ -0,0 +1 @@
+"""Prompt benchmarking system for evaluating LLMs via vLLM."""

prompt_bench/containers/finetune.py

@@ -0,0 +1,165 @@
+"""Docker container lifecycle management for Unsloth fine-tuning."""
+
+from __future__ import annotations
+
+import logging
+import subprocess
+from pathlib import Path
+from typing import Annotated
+
+import typer
+
+from python.prompt_bench.containers.lib import check_gpu_free
+
+logger = logging.getLogger(__name__)
+
+CONTAINER_NAME = "bill-finetune"
+FINETUNE_IMAGE = "bill-finetune:latest"
+DOCKERFILE_PATH = "/home/richie/dotfiles/python/prompt_bench/Dockerfile.finetune"
+DEFAULT_HF_CACHE = Path("/zfs/models/hf")
+
+
+def build_image() -> None:
+    """Build the fine-tuning Docker image."""
+    logger.info("Building fine-tuning image: %s", FINETUNE_IMAGE)
+    result = subprocess.run(
+        ["docker", "build", "-f", DOCKERFILE_PATH, "-t", FINETUNE_IMAGE, "."],
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        message = "Failed to build fine-tuning image"
+        raise RuntimeError(message)
+    logger.info("Image built: %s", FINETUNE_IMAGE)
+
+
+def start_finetune(
+    *,
+    dataset_path: Path,
+    output_dir: Path,
+    hf_cache: Path = DEFAULT_HF_CACHE,
+) -> None:
+    """Run the fine-tuning container.
+
+    Args:
+        dataset_path: Host path to the fine-tuning JSONL dataset.
+        output_dir: Host path where the trained model will be saved.
+        hf_cache: Host path to HuggingFace model cache (bind-mounted to avoid re-downloading).
+        validation_split: Fraction of data held out for validation.
+    """
+    dataset_path = dataset_path.resolve()
+    output_dir = output_dir.resolve()
+
+    if not dataset_path.is_file():
+        message = f"Dataset not found: {dataset_path}"
+        raise FileNotFoundError(message)
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    stop_finetune()
+
+    hf_cache = hf_cache.resolve()
+    hf_cache.mkdir(parents=True, exist_ok=True)
+
+    command = [
+        "docker",
+        "run",
+        "--name",
+        CONTAINER_NAME,
+        "--device=nvidia.com/gpu=all",
+        "--ipc=host",
+        "-v",
+        f"{hf_cache}:/root/.cache/huggingface",
+        "-v",
+        f"{output_dir}:/workspace/output/qwen-bill-summarizer",
+        "-v",
+        f"{dataset_path}:/workspace/dataset.jsonl:ro",
+        FINETUNE_IMAGE,
+        "--dataset",
+        "/workspace/dataset.jsonl",
+        "--output-dir",
+        "/workspace/output/qwen-bill-summarizer",
+    ]
+
+    logger.info("Starting fine-tuning container")
+    logger.info("  Dataset:    %s", dataset_path)
+    logger.info("  Output:     %s", output_dir)
+
+    result = subprocess.run(command, text=True, check=False)
+    if result.returncode != 0:
+        message = f"Fine-tuning container exited with code {result.returncode}"
+        raise RuntimeError(message)
+    logger.info("Fine-tuning complete. Model saved to %s", output_dir)
+
+
+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)
+
+
+def logs_finetune() -> str | None:
+    """Return recent logs from the fine-tuning container, or None if not running."""
+    result = subprocess.run(
+        ["docker", "logs", "--tail", "50", CONTAINER_NAME],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        return None
+    return result.stdout + result.stderr
+
+
+app = typer.Typer(help="Fine-tuning container management.")
+
+
+@app.command()
+def build() -> None:
+    """Build the fine-tuning Docker image."""
+    build_image()
+
+
+@app.command()
+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(
+        "/home/richie/dotfiles/data/output/qwen-bill-summarizer",
+    ),
+    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")
+    check_gpu_free()
+    start_finetune(
+        dataset_path=dataset,
+        output_dir=output_dir,
+        hf_cache=hf_cache,
+    )
+
+@app.command()
+def stop() -> None:
+    """Stop and remove the fine-tuning container."""
+    stop_finetune()
+
+
+@app.command()
+def logs() -> None:
+    """Show recent logs from the fine-tuning container."""
+    output = logs_finetune()
+    if output is None:
+        typer.echo("No running fine-tuning container found.")
+        raise typer.Exit(code=1)
+    typer.echo(output)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    app()
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/containers/lib.py

@@ -0,0 +1,23 @@
+from __future__ import annotations
+
+import logging
+import subprocess
+
+logger = logging.getLogger(__name__)
+
+
+def check_gpu_free() -> None:
+    """Warn if GPU-heavy processes (e.g. Ollama) are running."""
+    result = subprocess.run(
+        ["nvidia-smi", "--query-compute-apps=pid,process_name", "--format=csv,noheader"],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    if result.returncode != 0:
+        logger.warning("Could not query GPU processes: %s", result.stderr.strip())
+        return
+    processes = result.stdout.strip()
+    if processes:
+        logger.warning("GPU processes detected:\n%s", processes)
+        logger.warning("Consider stopping Ollama (sudo systemctl stop ollama) before benchmarking")

prompt_bench/containers/vllm.py

@@ -0,0 +1,70 @@
+"""Docker container lifecycle management for vLLM."""
+
+from __future__ import annotations
+
+import logging
+import subprocess
+
+logger = logging.getLogger(__name__)
+
+CONTAINER_NAME = "vllm-bench"
+VLLM_IMAGE = "vllm/vllm-openai:v0.19.0"
+
+
+def start_vllm(
+    *,
+    model: str,
+    port: int,
+    model_dir: str,
+    gpu_memory_utilization: float,
+) -> None:
+    """Start a vLLM container serving the given model.
+
+    Args:
+        model: HuggingFace model directory name (relative to model_dir).
+        port: Host port to bind.
+        model_dir: Host path containing HuggingFace model directories.
+        gpu_memory_utilization: Fraction of GPU memory to use (0-1).
+    """
+    command = [
+        "docker",
+        "run",
+        "-d",
+        "--name",
+        CONTAINER_NAME,
+        "--device=nvidia.com/gpu=all",
+        "--ipc=host",
+        "-v",
+        f"{model_dir}:/models",
+        "-p",
+        f"{port}:8000",
+        VLLM_IMAGE,
+        "--model",
+        f"/models/{model}",
+        "--served-model-name",
+        model,
+        "--gpu-memory-utilization",
+        str(gpu_memory_utilization),
+        "--max-model-len",
+        "4096",
+    ]
+    logger.info("Starting vLLM container with model: %s", model)
+    stop_vllm()
+    result = subprocess.run(command, capture_output=True, text=True, check=False)
+    if result.returncode != 0:
+        msg = f"Failed to start vLLM container: {result.stderr.strip()}"
+        raise RuntimeError(msg)
+    logger.info("vLLM container started: %s", result.stdout.strip()[:12])
+
+
+def stop_vllm() -> None:
+    """Stop and remove the vLLM benchmark container."""
+    logger.info("Stopping vLLM 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", "network", "disconnect", "-f", "bridge", CONTAINER_NAME],
+        capture_output=True,
+        check=False,
+    )
+    logger.info("vLLM container stopped and removed")

prompt_bench/downloader.py

@@ -0,0 +1,75 @@
+"""HuggingFace model downloader."""
+
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Annotated
+
+import typer
+from huggingface_hub import snapshot_download
+
+from python.prompt_bench.models import BenchmarkConfig
+
+logger = logging.getLogger(__name__)
+
+
+def local_model_path(repo: str, model_dir: str) -> Path:
+    """Return the local directory path for a HuggingFace repo."""
+    return Path(model_dir) / repo
+
+
+def is_model_present(repo: str, model_dir: str) -> bool:
+    """Check if a model has already been downloaded."""
+    path = local_model_path(repo, model_dir)
+    return path.exists() and any(path.iterdir())
+
+
+def download_model(repo: str, model_dir: str) -> Path:
+    """Download a HuggingFace model to the local model directory.
+
+    Skips the download if the model directory already exists and contains files.
+    """
+    local_path = local_model_path(repo, model_dir)
+
+    if is_model_present(repo, model_dir):
+        logger.info("Model already exists: %s", local_path)
+        return local_path
+
+    logger.info("Downloading model: %s -> %s", repo, local_path)
+    snapshot_download(
+        repo_id=repo,
+        local_dir=str(local_path),
+    )
+    logger.info("Download complete: %s", repo)
+    return local_path
+
+
+def download_all(config: BenchmarkConfig) -> None:
+    """Download every model listed in the config, top to bottom."""
+    for repo in config.models:
+        download_model(repo, config.model_dir)
+
+
+def main(
+    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")
+
+    if not config.is_file():
+        message = f"Config file does not exist: {config}"
+        raise typer.BadParameter(message)
+
+    benchmark_config = BenchmarkConfig.from_toml(config)
+    download_all(benchmark_config)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/finetune.py

@@ -0,0 +1,214 @@
+"""Fine-tune Qwen 3.5 4B on bill summarization data using Unsloth.
+
+Loads a ChatML-style JSONL dataset (system/user/assistant messages),
+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 \
+        --dataset output/finetune_dataset.jsonl \
+        --output-dir output/qwen-bill-summarizer
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Annotated
+
+import tomllib
+import typer
+from unsloth import FastLanguageModel
+from datasets import Dataset
+from transformers import TrainingArguments
+from trl import SFTTrainer
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class LoraConfig:
+    """LoRA adapter hyperparameters."""
+
+    rank: int
+    alpha: int
+    dropout: float
+    targets: list[str]
+
+
+@dataclass
+class TrainingConfig:
+    """Training loop hyperparameters."""
+
+    learning_rate: float
+    epochs: int
+    batch_size: int
+    gradient_accumulation: int
+    max_seq_length: int
+    warmup_ratio: float
+    weight_decay: float
+    logging_steps: int
+    save_steps: int
+
+
+@dataclass
+class FinetuneConfig:
+    """Top-level finetune configuration."""
+
+    base_model: str
+    lora: LoraConfig
+    training: TrainingConfig
+
+    @classmethod
+    def from_toml(cls, config_path: Path) -> FinetuneConfig:
+        """Load finetune config from a TOML file."""
+        raw = tomllib.loads(config_path.read_text())["finetune"]
+        return cls(
+            base_model=raw["base_model"],
+            lora=LoraConfig(**raw["lora"]),
+            training=TrainingConfig(**raw["training"]),
+        )
+
+
+def _messages_to_chatml(messages: list[dict]) -> str:
+    r"""Convert a message list to Qwen ChatML format.
+
+    Produces:
+        <|im_start|>system\n...\n<|im_end|>
+        <|im_start|>user\n...\n<|im_end|>
+        <|im_start|>assistant\n...\n<|im_end|>
+    """
+    parts = []
+    for message in messages:
+        role = message["role"]
+        content = message["content"]
+        parts.append(f"<|im_start|>{role}\n{content}<|im_end|>")
+    return "\n".join(parts)
+
+
+def load_dataset_from_jsonl(path: Path) -> Dataset:
+    """Load a ChatML JSONL file into a HuggingFace Dataset.
+
+    Each line must have {"messages": [{"role": ..., "content": ...}, ...]}.
+    Pre-formats into a `text` column with the Qwen ChatML template applied,
+    which SFTTrainer consumes directly.
+    """
+    records = []
+    with path.open(encoding="utf-8") as handle:
+        for raw_line in handle:
+            stripped = raw_line.strip()
+            if stripped:
+                entry = json.loads(stripped)
+                records.append({"text": _messages_to_chatml(entry["messages"])})
+    logger.info("Loaded %d examples from %s", len(records), path)
+    return Dataset.from_list(records)
+
+
+def main(
+    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,
+    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,
+) -> 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")
+
+    if not dataset_path.is_file():
+        message = f"Dataset not found: {dataset_path}"
+        raise typer.BadParameter(message)
+
+    config = FinetuneConfig.from_toml(config_path)
+
+    logger.info("Loading base model: %s", config.base_model)
+    model, tokenizer = FastLanguageModel.from_pretrained(
+        model_name=config.base_model,
+        max_seq_length=config.training.max_seq_length,
+        load_in_4bit=True,
+        dtype=None,
+    )
+
+    logger.info("Applying LoRA (rank=%d, alpha=%d)", config.lora.rank, config.lora.alpha)
+    model = FastLanguageModel.get_peft_model(
+        model,
+        r=config.lora.rank,
+        lora_alpha=config.lora.alpha,
+        lora_dropout=config.lora.dropout,
+        target_modules=config.lora.targets,
+        bias="none",
+        use_gradient_checkpointing="unsloth",
+        random_state=42,
+    )
+
+    full_dataset = load_dataset_from_jsonl(dataset_path)
+    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))
+    training_args = TrainingArguments(
+        output_dir=str(output_dir / "checkpoints"),
+        num_train_epochs=config.training.epochs,
+        per_device_train_batch_size=config.training.batch_size,
+        gradient_accumulation_steps=config.training.gradient_accumulation,
+        learning_rate=config.training.learning_rate,
+        warmup_ratio=config.training.warmup_ratio,
+        weight_decay=config.training.weight_decay,
+        lr_scheduler_type="cosine",
+        logging_steps=config.training.logging_steps,
+        save_steps=config.training.save_steps,
+        save_total_limit=3,
+        eval_strategy="steps",
+        eval_steps=config.training.save_steps,
+        load_best_model_at_end=True,
+        bf16=True,
+        optim="adamw_8bit",
+        seed=42,
+        report_to="none",
+    )
+
+    trainer = SFTTrainer(
+        model=model,
+        tokenizer=tokenizer,
+        train_dataset=train_dataset,
+        eval_dataset=validation_dataset,
+        args=training_args,
+        max_seq_length=config.training.max_seq_length,
+        packing=True,
+    )
+
+    logger.info(
+        "Starting training: %d train, %d val, %d epochs",
+        len(train_dataset),
+        len(validation_dataset),
+        config.training.epochs,
+    )
+    trainer.train()
+
+    merged_path = str(output_dir / "merged")
+    logger.info("Saving merged model to %s", merged_path)
+    model.save_pretrained_merged(merged_path, tokenizer, save_method="merged_16bit")
+
+    if save_gguf:
+        gguf_path = str(output_dir / "gguf")
+        logger.info("Saving GGUF to %s", gguf_path)
+        model.save_pretrained_gguf(gguf_path, tokenizer, quantization_method="q4_k_m")
+
+    logger.info("Done! Model saved to %s", output_dir)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/input/1.txt

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

prompt_bench/input/2.txt

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

prompt_bench/input/3.txt

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

prompt_bench/input/4.txt

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

prompt_bench/main.py

@@ -0,0 +1,215 @@
+"""CLI entry point for the prompt benchmarking system."""
+
+from __future__ import annotations
+
+import json
+import logging
+import time
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from pathlib import Path
+from typing import Annotated
+
+import typer
+
+from python.prompt_bench.containers.lib import check_gpu_free
+from python.prompt_bench.containers.vllm import start_vllm, stop_vllm
+from python.prompt_bench.downloader import is_model_present
+from python.prompt_bench.models import BenchmarkConfig
+from python.prompt_bench.vllm_client import VLLMClient
+
+logger = logging.getLogger(__name__)
+
+
+def discover_prompts(input_dir: Path) -> list[Path]:
+    """Find all .txt files in the input directory."""
+    prompts = list(input_dir.glob("*.txt"))
+    if not prompts:
+        message = f"No .txt files found in {input_dir}"
+        raise FileNotFoundError(message)
+    return prompts
+
+
+def _run_prompt(
+    client: VLLMClient,
+    prompt_path: Path,
+    *,
+    repo: str,
+    model_dir_name: str,
+    model_output: Path,
+    temperature: float,
+) -> tuple[bool, float]:
+    """Run a single prompt. Returns (success, elapsed_seconds)."""
+    filename = prompt_path.name
+    output_path = model_output / filename
+    start = time.monotonic()
+    try:
+        prompt_text = prompt_path.read_text()
+        response = client.complete(prompt_text, model_dir_name, temperature=temperature)
+        output_path.write_text(response)
+        elapsed = time.monotonic() - start
+        logger.info("Completed: %s / %s in %.2fs", repo, filename, elapsed)
+    except Exception:
+        elapsed = time.monotonic() - start
+        error_path = model_output / f"{filename}.error"
+        logger.exception("Failed: %s / %s after %.2fs", repo, filename, elapsed)
+        error_path.write_text(f"Error processing {filename}")
+        return False, elapsed
+    return True, elapsed
+
+
+def benchmark_model(
+    client: VLLMClient,
+    prompts: list[Path],
+    *,
+    repo: str,
+    model_dir_name: str,
+    model_output: Path,
+    temperature: float,
+    concurrency: int,
+) -> tuple[int, int]:
+    """Run all prompts against a single model in parallel.
+
+    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()]
+    skipped = len(prompts) - len(pending)
+    if skipped:
+        logger.info("Skipping %d prompts with existing output for %s", skipped, repo)
+
+    if not pending:
+        logger.info("Nothing to do for %s", repo)
+        return 0, 0
+
+    completed = 0
+    failed = 0
+    latencies: list[float] = []
+
+    wall_start = time.monotonic()
+    with ThreadPoolExecutor(max_workers=concurrency) as executor:
+        futures = [
+            executor.submit(
+                _run_prompt,
+                client,
+                prompt_path,
+                repo=repo,
+                model_dir_name=model_dir_name,
+                model_output=model_output,
+                temperature=temperature,
+            )
+            for prompt_path in pending
+        ]
+        for future in as_completed(futures):
+            success, elapsed = future.result()
+            latencies.append(elapsed)
+            if success:
+                completed += 1
+            else:
+                failed += 1
+    wall_elapsed = time.monotonic() - wall_start
+
+    attempted = completed + failed
+    avg_latency = sum(latencies) / attempted
+    throughput = attempted / wall_elapsed if wall_elapsed > 0 else 0.0
+    timing = {
+        "repo": repo,
+        "wall_seconds": wall_elapsed,
+        "attempted": attempted,
+        "completed": completed,
+        "failed": failed,
+        "avg_latency_seconds": avg_latency,
+        "throughput_prompts_per_second": throughput,
+        "concurrency": concurrency,
+    }
+    timing_path = model_output / "_timing.json"
+    timing_path.write_text(json.dumps(timing, indent=2))
+
+    return completed, failed
+
+
+def run_benchmark(
+    config: BenchmarkConfig,
+    input_dir: Path,
+    output_dir: Path,
+) -> None:
+    """Execute the benchmark across all models and prompts."""
+    prompts = discover_prompts(input_dir)
+    logger.info("Found %d prompts in %s", len(prompts), input_dir)
+
+    check_gpu_free()
+
+    total_completed = 0
+    total_failed = 0
+
+    for repo in config.models:
+        if not is_model_present(repo, config.model_dir):
+            logger.warning("Skipping (not downloaded): %s", repo)
+            continue
+
+        model_output = output_dir / repo
+        model_output.mkdir(parents=True, exist_ok=True)
+
+        logger.info("=== Benchmarking model: %s ===", repo)
+
+        stop_vllm()
+        try:
+            start_vllm(
+                model=repo,
+                port=config.port,
+                model_dir=config.model_dir,
+                gpu_memory_utilization=config.gpu_memory_utilization,
+            )
+        except RuntimeError:
+            logger.exception("Failed to start vLLM for %s, skipping", repo)
+            continue
+        logger.info("vLLM started for %s", repo)
+        try:
+            with VLLMClient(port=config.port, timeout=config.timeout) as client:
+                client.wait_ready(max_wait=config.vllm_startup_timeout)
+                completed, failed = benchmark_model(
+                    client,
+                    prompts,
+                    repo=repo,
+                    model_dir_name=repo,
+                    model_output=model_output,
+                    temperature=config.temperature,
+                    concurrency=config.concurrency,
+                )
+                total_completed += completed
+                total_failed += failed
+        finally:
+            stop_vllm()
+
+    logger.info("=== Benchmark complete ===")
+    logger.info("Completed: %d | Failed: %d", total_completed, total_failed)
+
+
+def main(
+    input_dir: Annotated[Path, typer.Argument(help="Directory containing input .txt prompt files")],
+    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")
+
+    if not input_dir.is_dir():
+        message = f"Input directory does not exist: {input_dir}"
+        raise typer.BadParameter(message)
+    if not config.is_file():
+        message = f"Config file does not exist: {config}"
+        raise typer.BadParameter(message)
+
+    benchmark_config = BenchmarkConfig.from_toml(config)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    run_benchmark(benchmark_config, input_dir, output_dir)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/models.py

@@ -0,0 +1,30 @@
+"""Pydantic models for benchmark configuration."""
+
+from __future__ import annotations
+
+import tomllib
+from typing import TYPE_CHECKING
+
+from pydantic import BaseModel
+
+if TYPE_CHECKING:
+    from pathlib import Path
+
+
+class BenchmarkConfig(BaseModel):
+    """Top-level benchmark configuration loaded from TOML."""
+
+    models: list[str]
+    model_dir: str = "/zfs/models/hf"
+    port: int = 8000
+    gpu_memory_utilization: float = 0.90
+    temperature: float = 0.0
+    timeout: int = 300
+    concurrency: int = 4
+    vllm_startup_timeout: int = 900
+
+    @classmethod
+    def from_toml(cls, config_path: Path) -> BenchmarkConfig:
+        """Load benchmark config from a TOML file."""
+        raw = tomllib.loads(config_path.read_text())["bench"]
+        return cls(**raw)

prompt_bench/summarization_prompts.py

@@ -0,0 +1,34 @@
+SUMMARIZATION_SYSTEM_PROMPT = """You are a legislative analyst extracting policy substance from Congressional bill text.
+
+Your job is to compress a bill into a dense, neutral structured summary that captures every distinct policy action — including secondary effects that might be buried in subsections.
+
+EXTRACTION RULES:
+- IGNORE: whereas clauses, congressional findings that are purely political statements, recitals, preambles, citations of existing law by number alone, and procedural boilerplate.
+- FOCUS ON: operative verbs — what the bill SHALL do, PROHIBIT, REQUIRE, AUTHORIZE, AMEND, APPROPRIATE, or ESTABLISH.
+- SURFACE ALL THREADS: If the bill touches multiple policy areas, list each thread separately. Do not collapse them.
+- BE CONCRETE: Name the affected population, the mechanism, and the direction (expands/restricts/maintains).
+- STAY NEUTRAL: No political framing. Describe what the text does, not what its sponsors claim it does.
+
+OUTPUT FORMAT — plain structured text, not JSON:
+
+OPERATIVE ACTIONS:
+[Numbered list of what the bill actually does, one action per line, max 20 words each]
+
+AFFECTED POPULATIONS:
+[Who gains something, who loses something, or whose behavior is regulated]
+
+MECHANISMS:
+[How it works: new funding, mandate, prohibition, amendment to existing statute, grant program, study commission, etc.]
+
+POLICY THREADS:
+[List each distinct policy domain this bill touches, even minor ones. Use plain language, not domain codes.]
+
+SYMBOLIC/PROCEDURAL ONLY:
+[Yes or No — is this bill primarily a resolution, designation, or awareness declaration with no operative effect?]
+
+LENGTH TARGET: 150-250 words total. Be ruthless about cutting. Density over completeness."""
+
+SUMMARIZATION_USER_TEMPLATE = """Summarize the following Congressional bill according to your instructions.
+
+BILL TEXT:
+{text_content}"""

prompt_bench/tools/build_finetune_dataset.py

@@ -0,0 +1,114 @@
+"""Build a fine-tuning JSONL dataset from batch request + output files.
+
+Joins the original request JSONL (system + user messages) with the batch
+output JSONL (assistant completions) by custom_id to produce a ChatML-style
+messages JSONL suitable for fine-tuning.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from pathlib import Path
+from typing import Annotated
+
+import typer
+
+logger = logging.getLogger(__name__)
+
+HTTP_OK = 200
+
+
+def load_requests(path: Path) -> dict[str, list[dict]]:
+    """Parse request JSONL into {custom_id: messages}."""
+    results: dict[str, list[dict]] = {}
+    with path.open(encoding="utf-8") as handle:
+        for raw_line in handle:
+            stripped = raw_line.strip()
+            if not stripped:
+                continue
+            record = json.loads(stripped)
+            custom_id = record["custom_id"]
+            messages = record["body"]["messages"]
+            results[custom_id] = messages
+    return results
+
+
+def load_completions(path: Path) -> dict[str, str]:
+    """Parse batch output JSONL into {custom_id: assistant_content}."""
+    results: dict[str, str] = {}
+    with path.open(encoding="utf-8") as handle:
+        for line_number, raw_line in enumerate(handle, 1):
+            stripped = raw_line.strip()
+            if not stripped:
+                continue
+            record = json.loads(stripped)
+            custom_id = record["custom_id"]
+            response = record.get("response", {})
+            if response.get("status_code") != HTTP_OK:
+                logger.warning("Skipping %s (line %d): status %s", custom_id, line_number, response.get("status_code"))
+                continue
+            body = response.get("body", {})
+            choices = body.get("choices", [])
+            if not choices:
+                logger.warning("Skipping %s (line %d): no choices", custom_id, line_number)
+                continue
+            content = choices[0].get("message", {}).get("content", "")
+            if not content:
+                logger.warning("Skipping %s (line %d): empty content", custom_id, line_number)
+                continue
+            results[custom_id] = content
+    return results
+
+
+def main(
+    requests_path: Annotated[Path, typer.Option("--requests", help="Batch request JSONL")] = Path(
+        "output/openai_batch/requests.jsonl",
+    ),
+    batch_output: Annotated[Path, typer.Option("--batch-output", help="Batch output JSONL")] = Path(
+        "batch_69d84558d91c819091d53f08d78f9fd6_output.jsonl",
+    ),
+    output_path: Annotated[Path, typer.Option("--output", help="Fine-tuning JSONL output")] = Path(
+        "output/finetune_dataset.jsonl",
+    ),
+    log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
+) -> None:
+    """Build fine-tuning dataset by joining request and output JSONL files."""
+    logging.basicConfig(level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
+
+    logger.info("Loading requests from %s", requests_path)
+    requests = load_requests(requests_path)
+    logger.info("Loaded %d requests", len(requests))
+
+    logger.info("Loading completions from %s", batch_output)
+    completions = load_completions(batch_output)
+    logger.info("Loaded %d completions", len(completions))
+
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    matched = 0
+    skipped = 0
+
+    with output_path.open("w", encoding="utf-8") as handle:
+        for custom_id, messages in requests.items():
+            assistant_content = completions.get(custom_id)
+            if assistant_content is None:
+                skipped += 1
+                continue
+
+            example = {
+                "messages": [*messages, {"role": "assistant", "content": assistant_content}],
+            }
+            handle.write(json.dumps(example, ensure_ascii=False))
+            handle.write("\n")
+            matched += 1
+
+    logger.info("Wrote %d examples to %s (skipped %d unmatched)", matched, output_path, skipped)
+
+
+def cli() -> None:
+    """Typer entry point."""
+    typer.run(main)
+
+
+if __name__ == "__main__":
+    cli()

prompt_bench/tools/count_tokens.py

@@ -0,0 +1,97 @@
+"""Sum token usage across compressed and uncompressed run directories."""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Annotated
+
+import typer
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class UsageTotals:
+    """Aggregate usage counters for a directory of run records."""
+
+    files: int = 0
+    errors: int = 0
+    prompt_tokens: int = 0
+    cached_tokens: int = 0
+    completion_tokens: int = 0
+    reasoning_tokens: int = 0
+    total_tokens: int = 0
+    per_file: list[tuple[str, int, int, int]] = field(default_factory=list)
+
+
+def tally_directory(directory: Path) -> UsageTotals:
+    """Return aggregated usage stats for every JSON record in a directory."""
+    totals = UsageTotals()
+    decoder = json.JSONDecoder()
+    for path in sorted(directory.glob("*.json")):
+        text = path.read_text().lstrip()
+        record, _ = decoder.raw_decode(text)
+        totals.files += 1
+        usage = record.get("usage")
+        if not usage:
+            totals.errors += 1
+            continue
+        prompt_tokens = usage.get("prompt_tokens", 0)
+        completion_tokens = usage.get("completion_tokens", 0)
+        total_tokens = usage.get("total_tokens", 0)
+        cached_tokens = (usage.get("prompt_tokens_details") or {}).get("cached_tokens", 0)
+        reasoning_tokens = (usage.get("completion_tokens_details") or {}).get("reasoning_tokens", 0)
+        totals.prompt_tokens += prompt_tokens
+        totals.completion_tokens += completion_tokens
+        totals.total_tokens += total_tokens
+        totals.cached_tokens += cached_tokens
+        totals.reasoning_tokens += reasoning_tokens
+        totals.per_file.append((path.name, prompt_tokens, completion_tokens, total_tokens))
+    return totals
+
+
+def log_totals(label: str, totals: UsageTotals) -> None:
+    """Log a one-block summary for a directory."""
+    counted = totals.files - totals.errors
+    average_total = totals.total_tokens / counted if counted else 0
+    logger.info("[%s]", label)
+    logger.info("  files          : %d (with usage: %d, errors: %d)", totals.files, counted, totals.errors)
+    logger.info("  prompt tokens  : %d", totals.prompt_tokens)
+    logger.info("  cached tokens  : %d", totals.cached_tokens)
+    logger.info("  completion tok : %d", totals.completion_tokens)
+    logger.info("  reasoning tok  : %d", totals.reasoning_tokens)
+    logger.info("  total tokens   : %d", totals.total_tokens)
+    logger.info("  avg total/file : %.1f", average_total)
+
+
+def main(
+    runs_dir: Annotated[Path, typer.Option("--runs-dir")] = Path("output/openai_runs_temp_1"),
+    log_level: Annotated[str, typer.Option("--log-level")] = "INFO",
+) -> None:
+    """Print token usage totals for the compressed and uncompressed run directories."""
+    logging.basicConfig(level=log_level, format="%(message)s")
+
+    grand = UsageTotals()
+    for label in ("compressed", "uncompressed"):
+        directory = runs_dir / label
+        if not directory.is_dir():
+            logger.warning("%s: directory not found at %s", label, directory)
+            continue
+        totals = tally_directory(directory)
+        log_totals(label, totals)
+        grand.files += totals.files
+        grand.errors += totals.errors
+        grand.prompt_tokens += totals.prompt_tokens
+        grand.cached_tokens += totals.cached_tokens
+        grand.completion_tokens += totals.completion_tokens
+        grand.reasoning_tokens += totals.reasoning_tokens
+        grand.total_tokens += totals.total_tokens
+
+    log_totals("grand total", grand)
+
+
+if __name__ == "__main__":
+    typer.run(main)

prompt_bench/vllm_client.py

@@ -0,0 +1,68 @@
+"""OpenAI-compatible client for vLLM's API."""
+
+from __future__ import annotations
+
+import logging
+import time
+from typing import Self
+
+import httpx
+
+logger = logging.getLogger(__name__)
+
+READY_POLL_INTERVAL = 2.0
+
+
+class VLLMClient:
+    """Talk to a vLLM server via its OpenAI-compatible API.
+
+    Args:
+        host: vLLM host.
+        port: vLLM port.
+        timeout: Per-request timeout in seconds.
+    """
+
+    def __init__(self, *, host: str = "localhost", port: int = 8000, timeout: int = 300) -> None:
+        """Create a client connected to a vLLM server."""
+        self._client = httpx.Client(base_url=f"http://{host}:{port}", timeout=timeout)
+
+    def wait_ready(self, max_wait: int) -> None:
+        """Poll /v1/models until the server is ready or timeout."""
+        deadline = time.monotonic() + max_wait
+        while time.monotonic() < deadline:
+            try:
+                response = self._client.get("/v1/models")
+                if response.is_success:
+                    logger.info("vLLM server is ready")
+                    return
+            except httpx.TransportError:
+                pass
+            time.sleep(READY_POLL_INTERVAL)
+        msg = f"vLLM server not ready after {max_wait}s"
+        raise TimeoutError(msg)
+
+    def complete(self, prompt: str, model: str, *, temperature: float = 0.0, max_tokens: int = 4096) -> str:
+        """Send a prompt to /v1/completions and return the response text."""
+        payload = {
+            "model": model,
+            "prompt": prompt,
+            "temperature": temperature,
+            "max_tokens": max_tokens,
+        }
+        logger.info("Sending prompt to %s (%d chars)", model, len(prompt))
+        response = self._client.post("/v1/completions", json=payload)
+        response.raise_for_status()
+        data = response.json()
+        return data["choices"][0]["text"]
+
+    def close(self) -> None:
+        """Close the HTTP client."""
+        self._client.close()
+
+    def __enter__(self) -> Self:
+        """Enter the context manager."""
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        """Close the HTTP client on exit."""
+        self.close()