Datasets:
Tasks:
Visual Question Answering
Languages:
English
Size:
100K<n<1M
ArXiv:
Tags:
vision-language-models
multimodal-reasoning
visual-reasoning
geometric-reasoning
spatial-reasoning
physical-reasoning
License:
File size: 13,450 Bytes
7cdd2f2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 | """Run fresh image-plus-question closed-loop queries through OpenRouter."""
from __future__ import annotations
import argparse
import asyncio
import json
import os
import random
import sys
from dataclasses import dataclass, field
from datetime import datetime, timezone
from pathlib import Path
from typing import Any
from dotenv import load_dotenv
from openai import APIConnectionError, APIStatusError, APITimeoutError, RateLimitError
REPO_ROOT = Path(__file__).resolve().parents[3]
TEST_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from grip_eval.config import MODELS, PRICE_TABLE_USD_PER_MILLION, validate_model_config
from grip_eval.manifest import load_manifest
from grip_eval.openrouter import OpenRouterClient, closed_loop_user_content
FORMAT_INSTRUCTION = "Answer with only the following, and nothing else:\n\nANSWER: <your answer>"
@dataclass
class CostState:
"""Concurrency-safe accumulated cost and stop state."""
total: float = 0.0
unknown_responses: int = 0
lock: asyncio.Lock = field(default_factory=asyncio.Lock)
stop: asyncio.Event = field(default_factory=asyncio.Event)
def utc_timestamp() -> str:
"""Return an ISO-8601 UTC timestamp."""
return datetime.now(timezone.utc).isoformat()
def completed_rows(path: Path) -> tuple[set[str], float, int]:
"""Read resumable row IDs and accumulated known cost from an existing JSONL."""
completed: set[str] = set()
cost = 0.0
unknown = 0
if not path.exists():
return completed, cost, unknown
with path.open("r", encoding="utf-8") as handle:
for line_number, line in enumerate(handle, start=1):
if not line.strip():
continue
try:
record = json.loads(line)
except json.JSONDecodeError as error:
raise ValueError(f"Invalid JSON in {path}:{line_number}: {error}") from error
row_id = str(record.get("row_id", ""))
if not row_id:
raise ValueError(f"Missing row_id in {path}:{line_number}")
# Failed requests remain pending so a corrected configuration can
# retry them. A later successful record with the same row_id wins
# during scoring.
response_text = str(record.get("response_raw", "")).strip()
if (
record.get("error")
or not response_text
or response_text.casefold() in {"none", "null"}
):
continue
completed.add(row_id)
if record.get("cost_usd") is None:
unknown += 1
else:
cost += float(record["cost_usd"])
return completed, cost, unknown
def should_retry(error: Exception) -> bool:
"""Return whether an OpenAI-compatible provider exception is transient."""
if isinstance(error, (RateLimitError, APIConnectionError, APITimeoutError)):
return True
return isinstance(error, APIStatusError) and error.status_code >= 500
async def append_jsonl(path: Path, record: dict[str, Any], lock: asyncio.Lock) -> None:
"""Append and flush one durable JSONL record."""
line = json.dumps(record, ensure_ascii=False, separators=(",", ":")) + "\n"
async with lock:
with path.open("a", encoding="utf-8", newline="") as handle:
handle.write(line)
handle.flush()
os.fsync(handle.fileno())
async def issue_query(
*,
client: OpenRouterClient,
model_key: str,
model_config: dict[str, Any],
row: dict[str, Any],
repository_root: Path,
retries: int,
backoff_seconds: float,
) -> dict[str, Any]:
"""Issue one fresh conversation with bounded transient-error retries."""
prompt_sent = f"{row['prompt']}\n\n{FORMAT_INSTRUCTION}"
image_path = repository_root / str(row["image_path"])
base = {
"row_id": row["row_id"],
"model": model_key,
"domain": row["domain"],
"stem": row["stem"],
"question_id": row["question_id"],
"level": int(row["level"]),
"image_path": row["image_path"],
"prompt_sent": prompt_sent,
}
if not image_path.is_file():
return {
**base,
"response_raw": "",
"input_tokens": 0,
"output_tokens": 0,
"reasoning_tokens": 0,
"cost_usd": None,
"latency_ms": 0,
"attempt": 0,
"error": f"image not found: {image_path}",
"timestamp": utc_timestamp(),
}
messages = [
{
"role": "user",
"content": closed_loop_user_content(image_path, prompt_sent),
}
]
for attempt in range(1, retries + 2):
try:
response = await client.chat(
model_key=model_key,
model_slug=model_config["slug"],
messages=messages,
max_tokens=int(model_config["max_tokens"]),
reasoning=model_config.get("reasoning"),
extra_body=model_config.get("extra_body"),
)
return {
**base,
"response_raw": response.text,
"input_tokens": response.input_tokens,
"output_tokens": response.output_tokens,
"reasoning_tokens": response.reasoning_tokens,
"cost_usd": response.cost_usd,
"latency_ms": response.latency_ms,
"attempt": attempt,
"error": None,
"timestamp": utc_timestamp(),
}
except Exception as error: # Provider exception classes vary by transport path.
transient = should_retry(error)
if transient and attempt <= retries:
delay = backoff_seconds * (2 ** (attempt - 1))
delay += random.uniform(0.0, min(1.0, delay * 0.1))
await asyncio.sleep(delay)
continue
return {
**base,
"response_raw": "",
"input_tokens": 0,
"output_tokens": 0,
"reasoning_tokens": 0,
"cost_usd": None,
"latency_ms": 0,
"attempt": attempt,
"error": f"{type(error).__name__}: {error}",
"timestamp": utc_timestamp(),
}
raise AssertionError("retry loop exited unexpectedly")
async def run_model(
*,
client: OpenRouterClient,
model_key: str,
rows: list[dict[str, Any]],
result_path: Path,
repository_root: Path,
concurrency: int,
retries: int,
backoff_seconds: float,
max_cost: float | None,
cost_state: CostState,
) -> int:
"""Run pending rows for one model with resumable append-only output."""
model_config = validate_model_config(model_key)
queue: asyncio.Queue[dict[str, Any]] = asyncio.Queue()
for row in rows:
queue.put_nowait(row)
file_lock = asyncio.Lock()
completed_now = 0
async def worker() -> None:
nonlocal completed_now
while not queue.empty() and not cost_state.stop.is_set():
async with cost_state.lock:
if max_cost is not None and cost_state.total >= max_cost:
cost_state.stop.set()
return
try:
row = queue.get_nowait()
except asyncio.QueueEmpty:
return
record = await issue_query(
client=client,
model_key=model_key,
model_config=model_config,
row=row,
repository_root=repository_root,
retries=retries,
backoff_seconds=backoff_seconds,
)
await append_jsonl(result_path, record, file_lock)
completed_now += 1
async with cost_state.lock:
if record["cost_usd"] is None:
cost_state.unknown_responses += 1
else:
cost_state.total += float(record["cost_usd"])
print(
f"{model_key}: {completed_now}/{len(rows)} | "
f"cost=${cost_state.total:,.4f} | "
f"unknown-cost={cost_state.unknown_responses}",
flush=True,
)
if max_cost is not None and cost_state.total >= max_cost:
cost_state.stop.set()
queue.task_done()
await asyncio.gather(*(worker() for _ in range(concurrency)))
return completed_now
def require_cost_fallback(model_keys: list[str], max_cost: float | None) -> None:
"""Require editable fallback prices when a hard cost ceiling is requested."""
if max_cost is None:
return
missing = [
key
for key in model_keys
if PRICE_TABLE_USD_PER_MILLION[key]["input"] is None
or PRICE_TABLE_USD_PER_MILLION[key]["output"] is None
]
if missing:
raise ValueError(
"--max-cost requires fallback input/output prices for every selected model; "
f"fill grip_eval/config.py for: {missing}"
)
async def async_main(args: argparse.Namespace) -> None:
"""Load the manifest, resume selected models, and close the provider client."""
load_dotenv()
api_key = os.getenv("OPENROUTER_API_KEY")
if not api_key:
raise EnvironmentError("OPENROUTER_API_KEY is not set")
# A paid run is intentionally restricted to one explicitly selected model.
model_keys = [args.model]
require_cost_fallback(model_keys, args.max_cost)
for key in model_keys:
validate_model_config(key)
repository_root = args.repository_root.resolve()
manifest = load_manifest(args.manifest)
args.results_dir.mkdir(parents=True, exist_ok=True)
pending_by_model: dict[str, list[dict[str, Any]]] = {}
accumulated_cost = 0.0
accumulated_unknown = 0
for model_key in model_keys:
result_path = args.results_dir / f"closed_loop_{model_key}.jsonl"
completed, prior_cost, prior_unknown = completed_rows(result_path)
pending = manifest.loc[~manifest["row_id"].isin(completed)]
if args.limit is not None:
pending = pending.head(args.limit)
pending_by_model[model_key] = pending.to_dict("records")
accumulated_cost += prior_cost
accumulated_unknown += prior_unknown
print(
f"{model_key}: {len(completed):,} existing, {len(pending):,} pending this run"
)
cost_state = CostState(total=accumulated_cost, unknown_responses=accumulated_unknown)
client = OpenRouterClient(api_key=api_key, timeout_seconds=args.timeout)
try:
for model_key in model_keys:
if cost_state.stop.is_set():
break
configured = int(MODELS[model_key]["default_concurrency"])
concurrency = args.concurrency or configured
await run_model(
client=client,
model_key=model_key,
rows=pending_by_model[model_key],
result_path=args.results_dir / f"closed_loop_{model_key}.jsonl",
repository_root=repository_root,
concurrency=concurrency,
retries=args.retries,
backoff_seconds=args.backoff,
max_cost=args.max_cost,
cost_state=cost_state,
)
finally:
await client.close()
print(
f"Run complete. Known cumulative cost=${cost_state.total:,.4f}; "
f"responses without cost={cost_state.unknown_responses}."
)
if cost_state.stop.is_set():
print("Stopped at the configured cost ceiling; unissued rows remain resumable.")
def parse_args() -> argparse.Namespace:
"""Parse command-line arguments."""
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--manifest", type=Path, default=TEST_ROOT / "plan" / "closed_loop_manifest.csv"
)
parser.add_argument("--results-dir", type=Path, default=TEST_ROOT / "results")
parser.add_argument("--repository-root", type=Path, default=REPO_ROOT)
parser.add_argument(
"--model",
required=True,
choices=list(MODELS),
help="Exactly one model key to run; there is deliberately no all-model default",
)
parser.add_argument("--limit", type=int, help="Maximum new rows per selected model")
parser.add_argument("--concurrency", type=int, help="Override per-model concurrency")
parser.add_argument("--max-cost", type=float, help="Cumulative USD ceiling")
parser.add_argument("--retries", type=int, default=5)
parser.add_argument("--backoff", type=float, default=1.0)
parser.add_argument("--timeout", type=float, default=120.0)
return parser.parse_args()
def main() -> None:
"""CLI entry point."""
args = parse_args()
if args.limit is not None and args.limit < 1:
raise ValueError("--limit must be positive")
if args.concurrency is not None and args.concurrency < 1:
raise ValueError("--concurrency must be positive")
if args.max_cost is not None and args.max_cost <= 0:
raise ValueError("--max-cost must be positive")
asyncio.run(async_main(args))
if __name__ == "__main__":
main()
|