hazyresearch / ama_prompting Goto Github PK

Dear authors,

I just want to know what is the definition of "good-quality" prompt and "imperfect" prompt. Are my understanding correct?

Good-quality prompt = prompt which mostly contains ground-truth question-answer pairs. (the answers are more tend to be correct)
imperfect prompt = prompt which contains many false question-answer pairs. (the answers are more likely to be incorrect)

Thanks,
Dylan

Hi,

This work is amazing, thanks for providing the code!
I have some questions about it.

1. In Line 47 of SST2_final.py, there is no ".sample(frac=1, random_state=0)", all demonstrations seem to be arranged by their labels in the few-shot baseline.
2. For NQ and WebQs, the evaluation metric is span overlap accuracy instead of exact matching, does GPT-3 use the same metric?
3. Can I change the order of demonstrations in the prompt chain? I tried it on WSC and found that the order can affect the evaluation results.

Thanks so much

I follow the instructions in readme, and meet this error. Do you know how to solve it?

ama_prompting/manifest/manifest/manifest.py", line 104, in init
raise ValueError(f"{list(kwargs.items())} arguments are not recognized.")
ValueError: [('session_id', None)] arguments are not recognized.

Hi, I ran

python3 tasks/drop_final.py \
  --run_zeroshot 1 \
  --run_fewshot 1 \
  --run_decomp 1 \
  --num_boost 3 \
  --k_shot 3 \
  --output_metrics_file ../ama_logs/metrics.json \
  --cache_connection ../ama_logs/manifest_cache.sqlite \
  --save_dir ../ama_logs/ama_final_runs

and I see:

Accuracy Zero Shot 0.10691393700228803
Accuracy Few Shot 0.21785939533525883
Accuracy by Boost Set Decomposed [0.1622917936425949, 0.16476214293285982, 0.17488727059013848]
Accuracy by Boost Set Decomposed Average 0.1673137357218644
Accuracy Boost Decomposed 0.16522288996755874
Saved metrics to ../ama_logs/metrics.json
Saved final data to ../ama_logs/ama_final_runs/drop

This seems lower than the numbers in the paper, is that expected?