NicoLab28 commited on
Commit
1fc0b40
·
verified ·
1 Parent(s): 78778ce

The plain matrices are not blind to what is stated once

Browse files
Files changed (1) hide show
  1. RELEASE_NOTES_v3.md +6 -4
RELEASE_NOTES_v3.md CHANGED
@@ -54,10 +54,12 @@ That costs 0.0027 of pure-text cosine on the residual and 0.0013 on the matrix
54
  which is why the 8B figures above are slightly below what a text-only corpus
55
  would have given.
56
 
57
- **Prefer the `-mlp` files.** The plain matrices are honest linear baselines and
58
- they are useful as such, but a linear map does not carry an attribute that the
59
- prompt states once. Ask for crossed ankles and a plain matrix will seat the
60
- subject cross-legged instead. The residual networks hold those.
 
 
61
 
62
  ## What this does not fix, and will not
63
 
 
54
  which is why the 8B figures above are slightly below what a text-only corpus
55
  would have given.
56
 
57
+ **Prefer the `-mlp` files.** They measure closer to the 32B, 0.9449 against
58
+ 0.9289 on the 8B, and on the demo scene both plain matrices get the seated pose
59
+ wrong: the prompt asks for one bare leg crossed over the other and they sit her
60
+ cross-legged. But the gap is narrower than that sounds — the cat, the straw hat,
61
+ the laundry and the bouncing knee all come through on the plain matrices too. On
62
+ this scene they missed one attribute among many, and it happened to be the pose.
63
 
64
  ## What this does not fix, and will not
65