Apply engram repair to weights: bit-identical outputs, trainable n-gram memory for downstream distill/SFT

README.md

@@ -237,3 +237,16 @@ Family repos: [Byrne-VE](https://huggingface.co/Quazim0t0/Byrne-VE) ·
 [Escarda-VE](https://huggingface.co/Quazim0t0/Escarda-VE) ·
 [Byrne-Docling-131M](https://huggingface.co/Quazim0t0/Byrne-Docling-131M) ·
 [Escarda-Docling-126M](https://huggingface.co/Quazim0t0/Escarda-Docling-126M)
+
+## Update: engram repair (behavior-preserving)
+
+The n-gram Engram memory in the original weights was degenerate: with the frozen
+LSH compressor at init scale, every token hashed to bucket 0, so only one table
+row ever received gradient. This revision rescales the (frozen) compressor and
+broadcasts the learned bucket-0 vector across all table rows.
+
+**Outputs are bit-identical to the previous revision** (verified: max logit
+difference 0.0 across a prompt battery). The only change: the Engram's hash now
+spreads across the full table and every bucket is independently trainable — so
+if you distill or SFT on top of this base, the n-gram memory will actually learn
+instead of staying a constant bias.

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9d8523ee2acea43ec5868d9e5b898c3e1774fe40f0ec4d56353c809f51056aa1
+oid sha256:c9d88f130b882daa2a64ca9cb070e5bc9de9128e801eec76b7846bee9f97cfb6
 size 389129376