Datasets:
File size: 3,359 Bytes
fa8b928 | 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 | #!/usr/bin/env python3
"""Generic OpenAI-chat OCR client for kbench engine comparison. Resumable (skips ids already in OUT).
Usage: engine_client.py PAGES_JSON OUT_JSONL URL MODEL PROMPT_KEY MAX_TOKENS(0=server default) WORKERS
Row out: {id, text, finish, error, seconds}. temperature 0, repetition_penalty 1.0, no logits processors."""
import base64, json, mimetypes, sys, time, urllib.request
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
PROMPTS = {
# baidu/Unlimited-OCR README single-image prompt (same string as our frozen eval)
"unlimited": "<image>document parsing.",
# deepseek-ai/DeepSeek-OCR README document prompt; <image> is inserted by the chat template in vLLM chat API
"deepseek": "<|grounding|>Convert the document to markdown. ",
# ATH-MaaS/OvisOCR2 README prompt
"ovis": ("Extract all readable content from the image in natural human reading order and output the result as a "
"single Markdown document. For charts or images, represent them using an HTML image tag: <img src=\"images/"
"bbox_{left}_{top}_{right}_{bottom}.jpg\" />, where left, top, right, bottom are bounding box coordinates "
"scaled to [0, 1000). Format formulas as LaTeX. Format tables as HTML: <table>...</table>. Transcribe all "
"other text as standard Markdown. Preserve the original text without translation or paraphrasing."),
}
pages_path, out_path, url, model, pkey, max_tokens, workers = sys.argv[1:8]
max_tokens, workers = int(max_tokens), int(workers); prompt = PROMPTS[pkey]
pages = json.load(open(pages_path))["pages"]
out = Path(out_path); done = {json.loads(l)["id"] for l in out.open()} if out.exists() else set()
todo = [p for p in pages if p["id"] not in done]
print(f"pages {len(pages)} done {len(done)} todo {len(todo)}", flush=True)
def one(p):
t = time.time(); img = Path(p["image"])
mime = mimetypes.guess_type(img.name)[0] or "image/png"
payload = {"model": model, "messages": [{"role": "user", "content": [
{"type": "image_url", "image_url": {"url": f"data:{mime};base64,{base64.b64encode(img.read_bytes()).decode()}"}},
{"type": "text", "text": prompt}]}],
"temperature": 0, "repetition_penalty": 1.0, "skip_special_tokens": False}
if max_tokens: payload["max_tokens"] = max_tokens
try:
req = urllib.request.Request(url, data=json.dumps(payload).encode(), headers={"Content-Type": "application/json"})
with urllib.request.urlopen(req, timeout=1200) as r: body = json.loads(r.read())
ch = body["choices"][0]
return {"id": p["id"], "text": ch["message"]["content"] or "", "finish": ch.get("finish_reason"), "error": None, "seconds": round(time.time() - t, 2)}
except Exception as e:
msg = repr(e)
if hasattr(e, "read"):
try: msg += " " + e.read().decode()[:300]
except Exception: pass
return {"id": p["id"], "text": "", "finish": None, "error": msg[:500], "seconds": round(time.time() - t, 2)}
with out.open("a") as f, ThreadPoolExecutor(workers) as ex:
for k, row in enumerate(ex.map(one, todo), 1):
f.write(json.dumps(row, ensure_ascii=False) + "\n"); f.flush()
if k % 200 == 0: print(k, row["id"], row["error"] or row["finish"], row["seconds"], flush=True)
print("CLIENT_DONE", flush=True)
|