IO-AIR commited on
Commit
23c6c03
·
verified ·
1 Parent(s): c1cb68e

KIP-Hash v3: the nano verify-and-dedup word (28 bits = 24 fuzzy bucket + 4 packed score; Spearman 0.719 vs teacher, fuzzy dup recall 0.958, 606 KB artifact, 224 docs/s stdlib)

Browse files
Files changed (1) hide show
  1. README.md +8 -0
README.md CHANGED
@@ -193,6 +193,14 @@ that reads only the stored word.
193
 
194
  ## Limitations (honest)
195
 
 
 
 
 
 
 
 
 
196
  - **Evaluation scale.** 157 verify rows, 96 dup pairs, 3604 negatives,
197
  14 adjudicated hard pairs — directional results; the staged scaled
198
  campaign (~2400 more labels across 8 domains) is one JWT away and
 
193
 
194
  ## Limitations (honest)
195
 
196
+ - **v2.0.1 note.** While building v3 we found that v2's projection
197
+ heads were fit on un-dequantized int8 projections (and its junk
198
+ thresholds were selected in a different score space than the module
199
+ compares): v2's `quality` (from the 128 bits) and all dedup numbers
200
+ are valid, but its deployed `quality_p`/`junk_prob` outputs are
201
+ mis-scaled until a v2.0.1 head re-fit. v3 is built on the corrected
202
+ dequantize-first pipeline end to end, and this note ships with v3 so
203
+ nobody deploys v2's projection heads blind.
204
  - **Evaluation scale.** 157 verify rows, 96 dup pairs, 3604 negatives,
205
  14 adjudicated hard pairs — directional results; the staged scaled
206
  campaign (~2400 more labels across 8 domains) is one JWT away and