#!/usr/bin/env python3 """ERQA with hosted VLM agents (OpenAI-compatible endpoints), one thread per agent, resumable. GROQ_API_KEY=... CEREBRAS_API_KEY=... GEMINI_API_KEY=... python scripts/erqa_api_agents.py [agent ...] Writes results/erqa/.jsonl with the raw reply and the extracted letter. """ import base64 # noqa: I001 import io import json import re import sys import threading import time from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "src")) from robotics.benchmarks import erqa # noqa: E402 from robotics.llm.providers import LLMError, chat # noqa: E402 AGENTS = { "gemini-3.1-flash-lite": ("gemini", "gemini-3.1-flash-lite", 4.5), "gemini-3.5-flash-lite": ("gemini", "gemini-3.5-flash-lite", 4.5), "gemma-4-31b": ("gemini", "gemma-4-31b-it", 4.5), "qwen3.8-27b@groq": ("groq", "qwen/qwen3.8-27b", 2.0), "qwen3.8-27b@cerebras": ("cerebras", "qwen-3.8-27b", 2.0), "gemini-robotics-er-2 (SOTA ref)": ("gemini", "gemini-robotics-er-2-preview", 6.0), } SUFFIX = "\nThink briefly, then give the final answer on the last line as 'Answer: '." def b64(img, max_side=640): im = img.copy() im.thumbnail((max_side, max_side)) buf = io.BytesIO() im.save(buf, "JPEG", quality=85) return "data:image/jpeg;base64," + base64.b64encode(buf.getvalue()).decode() def to_openai(content): out = [] for c in content: out.append({"type": "text", "text": c["text"]} if c["type"] == "text" else {"type": "image_url", "image_url": {"url": b64(c["image"])}}) return out def extract(text, opts): text = re.sub(r".*?", "", text or "", flags=re.S) m = re.findall(r"Answer\s*[::]\s*\(?([A-H])\b", text) if m and m[-1] in opts: return m[-1] m = re.findall(r"\b([A-H])\b", text[-80:]) for x in reversed(m): if x in opts: return x return None def run_agent(name, items, outdir): prov, model, pace = AGENTS[name] path = Path(outdir) / (re.sub(r"[^A-Za-z0-9.@-]+", "_", name) + ".jsonl") done = set() if path.exists(): done = {json.loads(line)["id"] for line in open(path) if json.loads(line)["pred"] is not None} for it in items: if it["id"] in done: continue msgs = [{"role": "user", "content": to_openai(erqa.interleave(it, SUFFIX))}] t = time.perf_counter() try: text, _ = chat(msgs, providers=[prov], model=model, max_tokens=8192, retries_per_provider=6, backoff=8.0, timeout=180) except LLMError as e: text = f"ERROR {str(e)[:150]}" pred = extract(text, it["options"]) with path.open("a") as f: f.write(json.dumps({"id": it["id"], "type": it["type"], "answer": it["answer"], "options": it["options"], "pred": pred, "agent": name, "latency_s": round(time.perf_counter() - t, 2), "raw": text[-600:]}) + "\n") time.sleep(pace) print("done", name, flush=True) def main(): names = sys.argv[1:] or [n for n in AGENTS if "SOTA" not in n] # SOTA ref: free daily quota exhausted items = erqa.load() import random random.Random(0).shuffle(items) # representative coverage at any point (file order clusters by type) Path("results/erqa").mkdir(parents=True, exist_ok=True) ths = [threading.Thread(target=run_agent, args=(n, items, "results/erqa")) for n in names] for t in ths: t.start() for t in ths: t.join() if __name__ == "__main__": main()