|
I'm now slowly accumulating game records. |
Answered by
Cryolite
Sep 12, 2023
Replies: 1 comment 2 replies
|
For me, 100Mil samples (not games) is enough to train a BERT model. |
2 replies
|
Thanks. Then how many Hanchans are approximately equivalent to 100M samples? |
|
On average, you can extract just under 150 samples per Hanchan. Therefore, by simple calculation, it would take about 7,000 Hanchan to gather 1M samples. However, since the samples extracted from a single Hanchan are highly correlated with each other, it would be better to collect samples from as many Hanchans as possible and then randomly sample the training examples from them. |
Answer selected by
FYNIXqwq
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
On average, you can extract just under 150 samples per Hanchan. Therefore, by simple calculation, it would take about 7,000 Hanchan to gather 1M samples. However, since the samples extracted from a single Hanchan are highly correlated with each other, it would be better to collect samples from as many Hanchans as possible and then randomly sample the training examples from them.