ArkidMitra commited on
Commit
19b930b
·
verified ·
1 Parent(s): cd6183f

v1.1: verified 213M fashion-vocab build (gate: 100% bit-identity on 18K texts, graceful OOV fallback)

Browse files
README.md CHANGED
@@ -35,7 +35,7 @@ model-index:
35
 
36
  # MODA Pro Lite
37
 
38
- **A 375M open-weights fashion encoder that beats [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m) — the strongest open fashion retrieval system — on catalog search, at full corpus, from one checkpoint with no serving machinery.**
39
 
40
  By [Hopit AI](https://hopit.ai). Weights are open (Apache-2.0); the training and data pipeline are proprietary.
41
 
@@ -84,6 +84,7 @@ Plain bi-encoder: one 768-d vector per image, one per query, cosine similarity.
84
  ## What's inside (summary)
85
 
86
  - **Base:** SigLIP2-base-384 (`timm/ViT-B-16-SigLIP2-384`), full fine-tune of both towers.
 
87
  - **Data:** 93K commercial product images (from [Marqo-GS-10M](https://huggingface.co/datasets/Marqo/marqo-GS-10M), Apache-2.0) re-annotated by a frontier VLM into structured attributes and four query registers, with graded relevance labels (99.9% coverage).
88
  - **Recipe:** graded contrastive learning over multiple query registers, an anchor loss to the frozen base to prevent catastrophic transfer loss, and weight-space interpolation for the released checkpoint. The training pipeline and dataset are not released.
89
  - **Distinctive property:** unlike typical fashion fine-tunes, Pro Lite does **not** collapse out of domain — the anchored recipe held generalist behavior while adding fashion skill (image-tower drift from base: 0.36%).
 
35
 
36
  # MODA Pro Lite
37
 
38
+ **A 213M open-weights fashion encoder that beats [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m) — the strongest open fashion retrieval system — on catalog search, at full corpus, from one checkpoint with no serving machinery.**
39
 
40
  By [Hopit AI](https://hopit.ai). Weights are open (Apache-2.0); the training and data pipeline are proprietary.
41
 
 
84
  ## What's inside (summary)
85
 
86
  - **Base:** SigLIP2-base-384 (`timm/ViT-B-16-SigLIP2-384`), full fine-tune of both towers.
87
+ - **Size — 213,159,938 params:** the base's 256K-token multilingual vocabulary is trimmed to the 44,646 pieces fashion text actually uses (BPE merge-closure preserved). Verified: **100.00% bit-identical** outputs on 12,000 benchmark queries and 6,000 held-out registers vs the untrimmed 375M model — every benchmark number in this card transfers exactly. Out-of-vocabulary input (emoji, non-Latin scripts, rare typos) re-segments gracefully into kept pieces (cosine 0.82–0.97 to full model), never a silent failure.
88
  - **Data:** 93K commercial product images (from [Marqo-GS-10M](https://huggingface.co/datasets/Marqo/marqo-GS-10M), Apache-2.0) re-annotated by a frontier VLM into structured attributes and four query registers, with graded relevance labels (99.9% coverage).
89
  - **Recipe:** graded contrastive learning over multiple query registers, an anchor loss to the frozen base to prevent catastrophic transfer loss, and weight-space interpolation for the released checkpoint. The training pipeline and dataset are not released.
90
  - **Distinctive property:** unlike typical fashion fine-tunes, Pro Lite does **not** collapse out of domain — the anchored recipe held generalist behavior while adding fashion skill (image-tower drift from base: 0.36%).
open_clip_config.json CHANGED
@@ -12,7 +12,7 @@
12
  },
13
  "text_cfg": {
14
  "context_length": 64,
15
- "vocab_size": 256000,
16
  "hf_tokenizer_name": "timm/ViT-B-16-SigLIP2-384",
17
  "tokenizer_kwargs": {
18
  "clean": "canonicalize"
 
12
  },
13
  "text_cfg": {
14
  "context_length": 64,
15
+ "vocab_size": 44646,
16
  "hf_tokenizer_name": "timm/ViT-B-16-SigLIP2-384",
17
  "tokenizer_kwargs": {
18
  "clean": "canonicalize"
open_clip_model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:50900c41094784fdd582a987609a4c89854d6fcd674273996f79f9f0efe46ce0
3
- size 1501954192
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:536c60b2f8dcb57492cd20b990edcfeba1c43bf47563f93945632b26f2a7c5b9
3
+ size 852674280
tokenizer.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:220c63d496e0c14e63eb656c91e0215e926202e4c74b1f089e09f1920d779b04
3
- size 34362885
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9c638d37b2bf0d43f97f3241e9d4ac2872485d485f79e981f89ef46764d583f2
3
+ size 3185164