Fork from Counting-Stars.
- Prepare a long context like
./harry_potter_all.txt - generate test data:
python gen_test_data_en.py- You may need to change prompt to get it right
- run evaluation:
python test -m <model_path> - insert result label into json file like:
"answer": "{\"little_penguin\": [1, 2, 3, 4]}",- three field was required in json file:
answer,sky_size,reference_counting_results
- three field was required in json file:
- generate visualized graph
Small model may not return the right answer like: [1, 2, 3, 4]. The viz_multi_page.py script will extract every number in the "answer" filed and reader each graph a single page in the pdf file.
- M: How many stars
- N: How long the context should be
- interval = MAX_LEN/N
- sky_size = [1 x interval, 2 x interval, ..., N x interval]
- each sky contain M stars