| import os,json,re,concurrent.futures as cf |
| from openai import OpenAI |
| from writing import W |
| c=OpenAI(base_url="https://router.huggingface.co/v1",api_key=os.environ["HF_TOKEN"],timeout=180) |
| P={w["id"]:w for w in W} |
| R={json.loads(l)["id"]:json.loads(l) for l in open('all_raw.jsonl') if '"write"' in l} |
| J=["google/gemma-3-27b-it","deepseek-ai/DeepSeek-V3.1"] |
| RUB="""You are a strict native Tamil-speaking evaluator. A user sent the PROMPT below to an AI model and got the RESPONSE. |
| Score the RESPONSE (integers 1-5): |
| - correctness: facts and advice are accurate and safe |
| - fluency: natural, grammatical Tamil (or natural Tanglish if the prompt was Tanglish); penalise literal translation, odd word choice, spelling errors |
| - instruction: follows the requested length/format/audience |
| - register_match: 1 if it replies in the same script/register as the prompt (Tamil script for Tamil prompts, Romanized Tamil for Tanglish prompts), else 0 |
| Return only JSON: {"correctness":n,"fluency":n,"instruction":n,"register_match":0or1,"note":"<=15 words"} |
| |
| PROMPT: |
| %s |
| |
| RESPONSE: |
| %s""" |
| def go(a): |
| wid,j=a |
| for t in range(4): |
| try: |
| r=c.chat.completions.create(model=j,messages=[{"role":"user","content":RUB%(P[wid]["prompt"],R[wid]["output"])}],max_tokens=300,temperature=0) |
| s=r.choices[0].message.content; d=json.loads(re.search(r"\{.*\}",s,re.S).group(0)); d.update(id=wid,judge=j); return d |
| except Exception as e: err=str(e) |
| return dict(id=wid,judge=j,error=err) |
| with cf.ThreadPoolExecutor(6) as ex: out=list(ex.map(go,[(w,j) for w in R for j in J])) |
| json.dump(out,open("judge.json","w"),ensure_ascii=False,indent=1) |
| for d in out: print(d) |
|
|
