ruoxi_sun
58657935fc
Vendor LLMmap / llm-verify / llm-fingerprint-detector under bash/fingerprint/tools so the three fingerprint benchmarks run with only /data1/eval mounted (no /data1/xii dependency): - run.py DEFAULT_TOOLS_ROOT prefers builtin tools/, falls back to /data1/xii - exclude .git / node_modules / template backups - detector dist/ (pre-built) retained; node_modules not needed at runtime
49 lines
1.8 KiB
Python
49 lines
1.8 KiB
Python
#!/usr/bin/env python
|
|
"""Non-interactive LLM fingerprinting helper.
|
|
|
|
Collect answers from a real target LLM for the 8 fingerprinting queries,
|
|
put one answer per line in a text file, then run:
|
|
|
|
python run_identify.py answers.txt [-k 6] [--model_path ./data/pretrained_models/default]
|
|
"""
|
|
import argparse
|
|
import os
|
|
|
|
from LLMmap.inference import load_LLMmap
|
|
|
|
|
|
def main():
|
|
ap = argparse.ArgumentParser(description='Fingerprint an LLM from a file of answers')
|
|
ap.add_argument('answers_file', type=str, help='Text file with one answer per line (8 lines)')
|
|
ap.add_argument('-k', type=int, default=6, help='Number of top candidates to print')
|
|
ap.add_argument('--model_path', type=str, default='./data/pretrained_models/default')
|
|
ap.add_argument('--device', type=str, default='cpu', choices=['cpu', 'cuda'])
|
|
ap.add_argument('--dump-queries', action='store_true',
|
|
help='Only print the fingerprinting queries, then exit')
|
|
args = ap.parse_args()
|
|
|
|
conf, llmmap = load_LLMmap(args.model_path, device=args.device)
|
|
|
|
if args.dump_queries:
|
|
print('Send these queries to the target LLM (one at a time) and save each response '
|
|
'on its own line in your answers file:\n')
|
|
for i, q in enumerate(llmmap.queries):
|
|
print(f'[{i + 1}] {q}\n')
|
|
return
|
|
|
|
with open(args.answers_file) as f:
|
|
answers = [line.rstrip('\n') for line in f if line.strip() != '']
|
|
|
|
if len(answers) != len(llmmap.queries):
|
|
raise SystemExit(
|
|
f'Expected {len(llmmap.queries)} answers (one per fingerprinting query), '
|
|
f'got {len(answers)}. Use --dump-queries to list the queries.'
|
|
)
|
|
|
|
print('### Predicted identity ###')
|
|
llmmap.print_result(llmmap(answers), k=args.k)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|