Update README.md
Browse files
README.md
CHANGED
@@ -63,7 +63,7 @@ tags:
|
|
63 |
|
64 |
#### 3. Run the model on a random sample of ~1k prompts on each of the 3 classes:
|
65 |
|
66 |
-
- It is important that the same `'pre-prompt x prompt'` sample be used with each ```"baseline"```, ```"undesired"
|
67 |
- This takes the total number of hidden-state samples I recorded to: ```3 x 10 x 1000 = 30,000``` (per layer!).
|
68 |
- This may seem like a lot compared to what other people are using to create control vectors with, but the theory regarding [estimation of covariance matrices](https://en.wikipedia.org/wiki/Estimation_of_covariance_matrices) shows we need at the ***very least*** a minimum of [one sample per feature](https://stats.stackexchange.com/questions/90045/how-many-samples-are-needed-to-estimate-a-p-dimensional-covariance-matrix) (and the models uploaded here have between 4k and 11.5k hidden state dimensions!).
|
69 |
|
|
|
63 |
|
64 |
#### 3. Run the model on a random sample of ~1k prompts on each of the 3 classes:
|
65 |
|
66 |
+
- It is important that the same `'pre-prompt x prompt'` sample be used with each (```"baseline"```, ```"undesired"```, ```"desired"```) triplet.
|
67 |
- This takes the total number of hidden-state samples I recorded to: ```3 x 10 x 1000 = 30,000``` (per layer!).
|
68 |
- This may seem like a lot compared to what other people are using to create control vectors with, but the theory regarding [estimation of covariance matrices](https://en.wikipedia.org/wiki/Estimation_of_covariance_matrices) shows we need at the ***very least*** a minimum of [one sample per feature](https://stats.stackexchange.com/questions/90045/how-many-samples-are-needed-to-estimate-a-p-dimensional-covariance-matrix) (and the models uploaded here have between 4k and 11.5k hidden state dimensions!).
|
69 |
|