ThakiCloud/Commerce-Review-KO-RoBERTa-Large
Token Classification • 0.3B • Updated • 13
Commerce AI Models
Online-shop task models. Korean: reviews. KO/EN: order-support agent. Click, repurchase, intent, search, rerank, attributes: not Korean-trained.
Note Review aspect + sentiment with verbatim evidence spans. Sealed test Joint F1 0.827 vs 0.682 (27B zero-shot); unsupported claims 6.1% vs 9.1%
ThakiCloud/Commerce-Slate-Click-FINN-LGBM
Tabular Classification • UpdatedNote Slate click (reward) model on FINN.no slates, Apache-2.0 data. Test MRR 0.463 vs 0.442 CTR baseline, 3/3 seeds CI>0
ThakiCloud/Commerce-Repurchase-OTTO-LGBM
Tabular Classification • UpdatedNote 7-day repurchase risk ranking on OTTO sessions, CC BY 4.0 data. AUC 0.692 vs 0.630 BG/NBD (cutoff 20, 3-seed ensemble), 9/9 cutoff-seed cells CI>0
ThakiCloud/Commerce-Rerank-Qwen3-8B
Text Ranking • 8B • Updated • 41Note Shopping query-product reranker, Qwen3-Reranker-8B LoRA merged to bf16, Marqo-GS-10M (Apache-2.0). Sealed test vs base zero-shot: +8.79pp novel_query, +8.47pp zero_shot (95% CI > 0)
ThakiCloud/Commerce-Search-EmbeddingGemma-2
Sentence Similarity • 0.7B • Updated • 43Note English product-search text embedding, EmbeddingGemma-2 text tower on ESCI (Apache-2.0). ESCI dev pool nDCG@10 0.626 vs 0.572 Qwen3-Embedding-8B (+5.4pp, CI>0); sealed test not used. Text-only
ThakiCloud/Commerce-Attribute-Qwen3.5-4B
Text Generation • Updated • 15Note Verbatim product-attribute extraction, Qwen3.5-4B LoRA on ABO (CC BY 4.0). Sealed test F1 0.928 vs 0.640 dictionary baseline, 0.469 base zero-shot; hallucination 3.4%
-
ThakiCloud/Commerce-Rerank-Qwen3-8B-LoRA
Text Ranking • Updated • 16
ThakiCloud/Commerce-OrderIntent-OTTO-LGBM
Tabular Classification • UpdatedNote 24-hour session order intent on OTTO, CC BY 4.0 data. Sealed test PR-AUC 0.413 vs 0.360 logistic regression (3-seed ensemble); dev beat or tied DCN-v2/DeepFM/GRU
ThakiCloud/Commerce-Agent-4B-v4.1
Text Generation • 4B • Updated • 66Note KO/EN order-support agent shipped with a mandatory Authorization Guard (package v4.1.1 adds stock/coupon feasibility checks). heldout_v41 (1,000 dialogues, 3 runs): 0 unauthorized state changes executed, approved changes 200/200, STRICT 1.000. Model alone not proven safe (1 unauthorized attempt in 1 run)
ThakiCloud/Commerce-SessionRec-OTTO-LGBM
Tabular Classification • UpdatedNote Next-click session recommender on OTTO (CC BY 4.0): co-visitation/repeat/item2vec candidates + LightGBM LambdaRank. Sealed test Click Recall@20 0.4926 vs 0.4701 rule baseline, Order Recall@20 0.6886 vs 0.6532 (paired 95% CI > 0); CPU only