Upload 7 files

.gitattributes

@@ -0,0 +1,4 @@
+models/mozilla/translations-en-fr-static/fp16.d384.npy filter=lfs diff=lfs merge=lfs -text
+models/mozilla/translations-en-fr-static/fp32.d384.npy filter=lfs diff=lfs merge=lfs -text
+models/mozilla/translations-en-fr-static/fp8_e4m3.d384.npy filter=lfs diff=lfs merge=lfs -text
+models/mozilla/translations-en-fr-static/fp8_e5m2.d384.npy filter=lfs diff=lfs merge=lfs -text

models/mozilla/translations-en-fr-static/README.md

@@ -0,0 +1,46 @@
+# Translations en-fr static embeddings (Firefox layout)
+
+Generated by `build_firefox_static_model.py` from a Model2Vec StaticModel.
+
+- source Model2Vec model: `en-fr-static-model2vec-sif`
+- embedding source: `bergamot_marian_wemb`
+- vocab size: 32000
+- dimensions exported: [384]
+- dtypes: ['fp32', 'fp16', 'fp8_e5m2', 'fp8_e4m3']
+
+## Weighting
+
+- scheme: `sif`
+- formula: `a / (a + probability)`
+- `a`: 0.0003
+
+The per-piece weighting is baked into the stored vectors, because Model2Vec
+pools with an unweighted mean. That is equivalent to weighting at pooling
+time: the two differ only by a positive scalar, which L2 normalization
+removes.
+
+## Common components (`ccr.d{dim}.npy`)
+
+Orthonormal directions that `StaticEmbeddingsPipeline` projects out of each
+pooled embedding before normalizing -- step (b) of SIF, which a bare
+mean-pool omits. Set `staticEmbeddingsOptions.commonComponents` to the number
+of rows to use; the file intentionally contains more than that so the value
+can be retuned without re-exporting.
+
+- fitted on ['text_en', 'text_fr'] from `['docs.parquet']`
+- fit by replicating the runtime path (tokenizer.json, gather, unweighted
+  mean, no normalization), so the distribution matches inference
+- re-fit per dimension; a truncated direction is not the principal direction
+  of the truncated space
+- fit on fp32 and reused for every dtype
+
+| dim | rows | variance explained (k=4) | (k=8) |
+|---|---|---|---|
+| 384 | 16 | 0.2003 | 0.2691 |
+
+## Caveats
+
+- The fitting corpus is a translated English browsing set, so the French side
+  is translationese rather than native French browsing.
+- `a` and the row count were selected on that same evaluation, so the
+  measured gain is mildly optimistic.

models/mozilla/translations-en-fr-static/fp16.d384.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b373ebddf5c799c14e604420cbf83ad6cf7d465070458076f6af5e131480865
+size 24576128

models/mozilla/translations-en-fr-static/fp32.d384.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6de82582b3f3e64c38fab6612ff4e7890d46a18ff543a2c0225c4ea0aef651f
+size 49152128

models/mozilla/translations-en-fr-static/fp8_e4m3.d384.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0bc3cce7c4f0b8a2495501ddd907f63050af8905a0c97d50d3203c758ef4880
+size 12288128

models/mozilla/translations-en-fr-static/fp8_e5m2.d384.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9425c9ce062360819e0ca6a5e146505c9943defde38ab20723ad064d78afd197
+size 12288128