meta-llama / llama3

The official Meta Llama 3 GitHub site
Other
21.42k stars 2.14k forks source link

The model is consistently modifying my numeric input #179

Open ill13 opened 1 month ago

ill13 commented 1 month ago

The issue is the model is consistently modifying my numeric input.

If you give the model a 4 digit numeric string like "8888", it will always change it to "88,888".
If you change your input to "7,777 + 3,333" the model consistently gives the correct answer.

I'm not looking for the model to do simple math, I'm looking for the model to not fuzz my input.

Trying What is 7777 + 3333? and both models keep changing my input to "77,777 + 33,333".

Both models will often give the correct answer to 77,777 + 33,333, however that's not what was asked.

This concern was confirmed by another user on Reddit.

Running:

subramen commented 2 weeks ago

Looks like you're using the quantized models, it might be hampering the model's performance on numerical data. I cannot replicate this issue on the official meta llama models, I get 11110 from both 8b and 70b models.

Try increasing the temperature, 0.01 sounds quite low.