File size: 1,765 Bytes
7f9a652
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Turn questions into prompts for any LLM / RAG / agent system.



usage: python make_prompts.py --qa data/dev --out prompts.jsonl [--with-context-hint]

Each output line: {"id", "prompt"}. Send `prompt` to your system, then write

{"id", "answer"} lines and score them with score.py.

"""
import argparse, glob, json, os

INSTR = ("Answer the question using only the named dataset. Reply with the answer only: "
         "a number (no units), a name or code exactly as written in the data, true/false, "
         "or a JSON list when a list is asked for. If the dataset does not contain the answer, reply null.")


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument('--qa', required=True)
    ap.add_argument('--out', required=True)
    ap.add_argument('--with-context-hint', action='store_true', help='add the Hugging Face repo id so a tool-using agent knows where to look')
    a = ap.parse_args()
    files = sorted(glob.glob(os.path.join(a.qa, '*.jsonl'))) if os.path.isdir(a.qa) else [a.qa]
    n = 0
    with open(a.out, 'w', encoding='utf-8') as f:
        for fn in files:
            for line in open(fn, encoding='utf-8'):
                it = json.loads(line)
                p = INSTR + '\n\n'
                if a.with_context_hint:
                    p += f"Dataset: https://huggingface.co/datasets/{it['dataset']} (revision {it['revision']})\n"
                if it.get('answer_format'):
                    p += f"Answer format: {it['answer_format']}\n"
                p += f"Question: {it['question']}"
                f.write(json.dumps({'id': it['id'], 'prompt': p}, ensure_ascii=False) + '\n')
                n += 1
    print(n, 'prompts ->', a.out)


if __name__ == '__main__':
    main()