Clarify read projection usage in README
Added note about read projection usage for exporting results.
This commit is contained in:
@@ -234,8 +234,12 @@ editing. A mode toggle switches between:
|
|||||||
- **Read projection** (pure-weights) — a change of basis of the downstream reads so
|
- **Read projection** (pure-weights) — a change of basis of the downstream reads so
|
||||||
the **live preview matches the exported checkpoint exactly**. Use this to save a model and preview the result.
|
the **live preview matches the exported checkpoint exactly**. Use this to save a model and preview the result.
|
||||||
|
|
||||||
|
|
||||||
|
#### Read projection is what you want to use if your intent is to export the result.
|
||||||
|
|
||||||

|

|
||||||
|
|
||||||
|
|
||||||
> **Architecture note**: models whose layers normalize their *writes* into the
|
> **Architecture note**: models whose layers normalize their *writes* into the
|
||||||
> residual stream (Gemma 2/3 style, `pre/post_feedforward_layernorm`) can't take
|
> residual stream (Gemma 2/3 style, `pre/post_feedforward_layernorm`) can't take
|
||||||
> the read projection. On those, the toggle offers **Global projection** (W_U
|
> the read projection. On those, the toggle offers **Global projection** (W_U
|
||||||
|
|||||||
Reference in New Issue
Block a user