ArkidMitra commited on
Commit
0199a3e
Β·
verified Β·
1 Parent(s): d9cb519

Add the organization card

Browse files
Files changed (1) hide show
  1. README.md +48 -111
README.md CHANGED
@@ -1,127 +1,64 @@
1
- # Hopit AI
2
-
3
- Fashion AI infrastructure β€” retrieval, trend intelligence and catalogue models for fashion
4
- commerce, measured in public against open benchmarks.
5
-
6
- ## The MODA family
7
-
8
- | model | what it is | availability |
9
- |---|---|---|
10
- | **[MODA Pro Lite+](https://huggingface.co/HopitAI/moda-pro-lite-plus)** (213M) | Trained fashion encoder with a calibrated serving recipe. Leads the ≀250M class on catalogue and title search. | open weights |
11
- | **[MODA Duo](https://huggingface.co/HopitAI/moda-duo)** | Routes each query to the constituent that suits its shape. Beats either alone on a mixed workload, at one encoder per search. | open recipe Β· open weights |
12
- | **[MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m)** (203M) | Zero-new-parameter serving recipe over frozen FashionSigLIP. Strongest of the open models on long descriptions. | open source Β· open weights |
13
- | **MODA Pro** | Hosted retrieval system. Rank 1 or 2 on five of six full-corpus benchmarks. | closed Β· [hopit.ai](https://hopit.ai) |
14
- | **[MODA-SigLIP-Distilled](https://huggingface.co/HopitAI/moda-fashion-distilled)** (203M) | Image-to-image specialist. #1 open model on LookBench. | open weights |
15
-
16
- ## Text-to-image retrieval
17
-
18
- MAP@10, full corpus, all ground-truth queries, one evaluator (`pytrec_eval map_cut.10`).
19
- Every system in each size class is shown β€” ours and everyone else's, wins and losses.
20
- Columns ordered by number of first places, then by top-two finishes.
21
-
22
- ### Models at 250M parameters or fewer
23
 
24
- | benchmark | [MODA Pro Lite+](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M | [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m)<br>203M | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP)<br>203M | [MODA Pro Lite](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M |
25
- |---|---:|---:|---:|---:|
26
- | KAGL | **0.3201** | 0.2887 | 0.2769 | 0.3055 |
27
- | Polyvore | **0.4049** | 0.3726 | 0.3664 | 0.3952 |
28
- | Atlas | **0.1904** | 0.1862 | 0.1826 | 0.1814 |
29
- | Fashion200K | 0.1846 | **0.1946** | 0.1858 | 0.1758 |
30
- | DeepFashion In-Shop | 0.1026 | **0.1642** | 0.1586 | 0.0930 |
31
- | DeepFashion Multimodal | 0.0133 | 0.0147 | **0.0148** | 0.0118 |
32
 
33
- ### All systems
34
 
35
- | benchmark | [SO400M](https://huggingface.co/timm/ViT-SO400M-14-SigLIP-384)<br>878M | MODA Pro<br>hosted | [MODA Duo](https://huggingface.co/HopitAI/moda-duo)<br>2 indexes | [MODA](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m)<br>203M | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP)<br>203M | [MODA Pro Lite+](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M | [MODA Pro Lite](https://huggingface.co/HopitAI/moda-pro-lite)<br>213M | [ZooClaw](https://huggingface.co/srpone/zooclaw-fashionsiglip2)<br>375M |
36
- |---|---:|---:|---:|---:|---:|---:|---:|---:|
37
- | KAGL | **0.3370** | 0.3221 | 0.3201 | 0.2887 | 0.2769 | 0.3201 | 0.3055 | 0.2951 |
38
- | Polyvore | **0.4378** | 0.4106 | 0.4049 | 0.3726 | 0.3664 | 0.4049 | 0.3952 | 0.3804 |
39
- | Atlas | **0.2309** | 0.2009 | 0.1904 | 0.1862 | 0.1826 | 0.1904 | 0.1814 | 0.1583 |
40
- | Fashion200K | 0.1353 | **0.2065** | 0.1866 | 0.1946 | 0.1858 | 0.1846 | 0.1758 | 0.1775 |
41
- | DeepFashion In-Shop | **0.1694** | 0.1663 | 0.1640 | 0.1642 | 0.1586 | 0.1026 | 0.0930 | 0.1024 |
42
- | DeepFashion Multimodal | 0.0079 | 0.0137 | **0.0159** | 0.0147 | 0.0148 | 0.0133 | 0.0118 | 0.0099 |
43
 
44
- ## Image-to-image retrieval
45
 
46
- Finding visually similar products from a photo, on LookBench (2,345 queries,
47
- [Gao et al., 2026](https://arxiv.org/abs/2601.14706)). Every system measured through one
48
- harness; FashionSigLIP was reproduced here before any comparison was made.
49
 
50
- | model | params | dim | Fine R@1 | nDCG@5 |
51
- |---|---:|---:|---:|---:|
52
- | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | 203M | 768 | **67.63** | 53.85 |
53
- | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) @256d | 203M | 256 | 67.42 | **57.48** |
54
- | GR-Pro (SerendipityOne, closed) | n/a | 1024 | 67.38 | β€” |
55
- | [moda-fashion-crossdomain](https://huggingface.co/HopitAI/moda-fashion-crossdomain) | 203M | 768 | 66.52 | 52.46 |
56
- | [Tianmu-MERE](https://huggingface.co/TianmuLab/Tianmu-MERE) (Kuaishou) | 1.24B | 512 | 65.99 | 52.37 |
57
- | [GR-Lite](https://huggingface.co/srpone/gr-lite) (SerendipityOne) | 303M | 1024 | 65.71 | β€” |
58
- | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP) (our reproduction) | 203M | 768 | 63.84 | 49.63 |
59
- | [FashionSigLIP](https://huggingface.co/Marqo/marqo-fashionSigLIP) (published) | 203M | 768 | 62.77 | 49.44 |
60
 
61
- **Tianmu-MERE**, the top open model on the LookBench multi-attribute leaderboard, was
62
- evaluated through this same harness and reproduces its own published subset scores within
63
- 0.5 points β€” which independently validates the harness. moda-fashion-distilled leads it on
64
- all four subsets, on Fine R@1 and nDCG@5 alike, at 6Γ— fewer parameters.
65
 
66
- The configuration above was chosen from eight candidates ranked on LookBench itself, so the
67
- headline carries the upward bias of a maximum over eight. It was independently reproduced
68
- in-harness at 67.59, and the runner-up configurations land within 0.3 points.
69
 
70
- ## Which model should I use?
 
 
 
 
71
 
72
- | your workload | use | why |
73
  |---|---|---|
74
- | Catalogue / title search β€” `red floral maxi dress` | [moda-pro-lite](https://huggingface.co/HopitAI/moda-pro-lite) | leads the ≀250M class on KAGL, Polyvore and Atlas |
75
- | Long descriptions, exact-item search | [moda-duo](https://huggingface.co/HopitAI/moda-duo) | routes long queries to the description-strong constituent |
76
- | A mixture of both | [moda-duo](https://huggingface.co/HopitAI/moda-duo) | beats either constituent alone on a mixed workload |
77
- | Agentic and multi-step search | MODA Pro ([hosted](https://hopit.ai)) | one API, no indexes to build or operate |
78
- | Visually similar products from a photo | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | #1 open model on LookBench |
79
- | Same, but index size matters | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) | truncatable embeddings, 64–768d |
80
- | Edge / mobile deployment | [moda-fashion-vision-fp16](https://huggingface.co/HopitAI/moda-fashion-vision-fp16) | 186 MB vision tower |
81
-
82
- ## Quick start
83
-
84
- ```bash
85
- pip install open_clip_torch pillow
86
- ```
87
-
88
- ```python
89
- import open_clip, torch
90
-
91
- model, _, preprocess = open_clip.create_model_and_transforms("hf-hub:HopitAI/moda-pro-lite")
92
- tokenizer = open_clip.get_tokenizer("hf-hub:HopitAI/moda-pro-lite")
93
- model.eval()
94
-
95
- with torch.no_grad():
96
- img = torch.nn.functional.normalize(model.encode_image(preprocess(image).unsqueeze(0)), dim=-1)
97
- txt = torch.nn.functional.normalize(model.encode_text(tokenizer(["red floral maxi dress"])), dim=-1)
98
- score = (txt @ img.T).item()
99
- ```
100
-
101
- Each model repository ships a runnable script β€” `serving_ann.py` for the retrieval systems
102
- (including the ANN index), `inference.py` for the image-to-image checkpoints.
103
-
104
 
105
- ## Models
 
 
106
 
107
- | model | task | params | best for |
108
- |---|---|---|---|
109
- | [moda-duo](https://huggingface.co/HopitAI/moda-duo) | text β†’ image | recipe over two encoders | mixed workloads; routes per query, one encoder per search |
110
- | [moda-pro-lite](https://huggingface.co/HopitAI/moda-pro-lite) | text β†’ image | 213M | catalogue and title search; ships its serving recipe |
111
- | [moda-fashionsiglip-multiview-203m](https://huggingface.co/HopitAI/moda-fashionsiglip-multiview-203m) | text β†’ image | 203M | long descriptions and exact-item retrieval |
112
- | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | image β†’ image | 203M | best overall visual similarity |
113
- | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) | image β†’ image | 203M | truncatable embeddings, 64–768d |
114
- | [moda-fashion-distilled-512d](https://huggingface.co/HopitAI/moda-fashion-distilled-512d) | image β†’ image | 203M | compact indexes |
115
- | [moda-fashion-crossdomain](https://huggingface.co/HopitAI/moda-fashion-crossdomain) | image β†’ image | 203M | street-to-shop |
116
- | [moda-fashion-vision-fp16](https://huggingface.co/HopitAI/moda-fashion-vision-fp16) | image β†’ image | 93M vision | edge and mobile, 186 MB |
117
 
118
- ## How these numbers are measured
119
 
120
- Full corpus, no subsampling β€” every query scored against the entire benchmark corpus
121
- (42,537–201,624 items). One evaluator for every system, including competitors. Serving
122
- recipes are selected on held-out development data and frozen in writing before any benchmark
123
- is scored. Losses are shown, not hidden.
 
124
 
125
- Receipts, per-query results and confidence intervals: **[github.com/hopit-ai/Moda](https://github.com/hopit-ai/Moda)**
126
- Β· Benchmarks: **[hopit-ai.github.io/Moda](https://hopit-ai.github.io/Moda/)**
127
- Β· Hosted retrieval and agentic search: **[hopit.ai](https://hopit.ai)**
 
1
+ ---
2
+ title: README
3
+ emoji: πŸ‘—
4
+ colorFrom: indigo
5
+ colorTo: gray
6
+ sdk: static
7
+ pinned: false
8
+ ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9
 
10
+ # Hopit AI
 
 
 
 
 
 
 
11
 
12
+ World models for fashion: search, retrieval, ranking, and multimodal reasoning.
13
 
14
+ We publish two tracks. Each one ships its weights, its evaluation protocol, and the
15
+ prediction files behind every number we quote, so our claims can be checked rather
16
+ than taken on trust.
 
 
 
 
 
17
 
18
+ ## Moda: search and retrieval
19
 
20
+ Fashion image encoders for search, recommendation, and visual similarity. The
21
+ distilled encoder is the one we serve, and it is the backbone under the attribute
22
+ models below.
23
 
24
+ | Model | What it is |
25
+ |---|---|
26
+ | [moda-fashion-distilled](https://huggingface.co/HopitAI/moda-fashion-distilled) | The served encoder. Start here. |
27
+ | [moda-fashion-matryoshka](https://huggingface.co/HopitAI/moda-fashion-matryoshka) | Nested embeddings, so one model serves several dimensions |
28
+ | [moda-fashion-distilled-512d](https://huggingface.co/HopitAI/moda-fashion-distilled-512d) | Smaller embeddings for tighter index budgets |
29
+ | [moda-fashion-crossdomain](https://huggingface.co/HopitAI/moda-fashion-crossdomain) | Street photos matched to catalogue product shots |
30
+ | [moda-fashion-vision-fp16](https://huggingface.co/HopitAI/moda-fashion-vision-fp16) | Half-precision vision tower for cheaper inference |
 
 
 
31
 
32
+ Benchmarks: [hopit-ai.github.io/Moda](https://hopit-ai.github.io/Moda/) Β·
33
+ Code: [github.com/hopit-ai/Moda](https://github.com/hopit-ai/Moda)
 
 
34
 
35
+ ## MODA_NER: attribute extraction
 
 
36
 
37
+ Turning a fashion image into structured product data: category, colour, fit,
38
+ neckline, sleeve length, pattern, material. Each model is scored on one frozen
39
+ track of the MODA General Attribute Suite, and the tracks are never averaged
40
+ together, because a model can be strong on clean product shots and weak on
41
+ full-body photos.
42
 
43
+ | Model | Input | Licence |
44
  |---|---|---|
45
+ | [moda-ner-v-crop](https://huggingface.co/HopitAI/moda-ner-v-crop) | A cropped garment | MIT |
46
+ | [moda-ner-v-catalog](https://huggingface.co/HopitAI/moda-ner-v-catalog) | A catalogue product image | CC BY-NC 4.0 |
47
+ | [moda-ner-v-fullbody](https://huggingface.co/HopitAI/moda-ner-v-fullbody) | A full-body photo | CC BY-NC 4.0 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48
 
49
+ Two of these are evaluated against research-only corpora whose terms reach derived
50
+ data, so their weights are non-commercial. That binds us too: those weights are not
51
+ part of our paid product.
52
 
53
+ Benchmarks: [hopit-ai.github.io/Moda_ner](https://hopit-ai.github.io/Moda_ner/) Β·
54
+ Code: [github.com/hopit-ai/Moda_ner](https://github.com/hopit-ai/Moda_ner)
 
 
 
 
 
 
 
 
55
 
56
+ ## How we evaluate
57
 
58
+ We freeze the test set before running a model, keep related images out of both
59
+ training and test data, hash every prediction file before scoring it, score every
60
+ row instead of quietly dropping the awkward ones, and require a real gain on each
61
+ track rather than a good average. We publish the losing runs alongside the winning
62
+ ones.
63
 
64
+ Contact: [hopit.ai](https://hopit.ai)