bvrtuu/trendyol-eticaret-2026-models

Trendyol E-Ticaret 2026 Shared Models

DEVAM EDEN İŞ (2026-07-07, public LB 0.927 ~12. sıra): Güncel yükseliş yeni MODEL eğitmiyor — API-hakem cerrahi-flip fabrikası (DeepSeek V3+R1 kesişimi) ile anchor üzerinde etiket-düzeltmesi yapıyor. Tüm devam malzemesi GitHub'da: docs/HANDOFF-judge-flip-factory.md (kurulum + komut sırası + gate'ler + bütçe). Bu HF repo'daki model ağırlıkları TABAN olarak duruyor, değişmedi. Q1 (Qwen3-Reranker-4B FT) gate'ten kaldı (veri-gürültüsü tavanı) — yüklenmedi.

Public model artifacts for the Trendyol E-Ticaret Yarismasi 2026 Kaggle pipeline.

Last verified Hub visibility: public on 2026-06-29.

Contents

These are the default checkpoints downloaded by scripts/download_models.py. The public Hub repo can also keep older experiment folders; use --models all only when those are needed.

Folder	Source model	Text mode	Max length	Notes
ce/	dbmdz/bert-base-turkish-cased	base	64	Public LB anchor around 0.80
ce_electra/	dbmdz/electra-base-turkish-cased-discriminator	base	64	Diversity CE, public LB around 0.80
ce_rich/	dbmdz/bert-base-turkish-cased	rich	128	Uses selected attributes/gender/age; public LB 0.79
ce_hardft_faiss10/	ce/ fine-tune	base	64	FAISS hard-negative fine-tune; ce_l3hard_pos0.240.csv reached Public LB 0.82
ce_hardft_deep/	ce/ fine-tune	base	64	Deeper FAISS hard-negative fine-tune with up to 24 negatives per term
ce_eldeep/	ce_electra/ fine-tune	base	64	ELECTRA version of the deeper FAISS hard-negative fine-tune
ce_xlmr/	FacebookAI/xlm-roberta-large (560M)	base	64	XLM-R large on deep negatives; ce_l3xlmr_pos0.240.csv reached Public LB 0.87 (trained bf16, grad-ckpt)
ce_bge/	BAAI/bge-reranker-v2-m3 (568M)	base	64	Reranker-pretrained on deep negatives; ce_l3bge_pos0.240.csv Public LB 0.87, proxy 0.6691
ce_bge_attr/	BAAI/bge-reranker-v2-m3 (568M)	attr	96	bge + curated attributes (attr text mode); ce_l3bgeattr_pos0.240.csv Public LB 0.88
ce_bgeattrmm/	BAAI/bge-reranker-v2-m3 (568M)	attr	96	+ deep+attrmm negatifler (cap 32), seed 42; tekil submit LB 0.88 (düz)
ce_bgeattrmm43/	BAAI/bge-reranker-v2-m3 (568M)	attr	96	aynı reçete seed 43; 0.89 ortalamasının üyesi
ce_bgeattrmmps/	BAAI/bge-reranker-v2-m3 (568M)	attr	96	+ pseudo_v2_strict.csv; 0.89 ortalamasının üyesi
ce_bgeattrctx/	BAAI/bge-reranker-v2-m3 (568M)	attrctx	128	co-candidate query context; tekil proxy düz, attr2ctx kombinasyon adayı
ce_bgeattrllm/	BAAI/bge-reranker-v2-m3 (568M)	attr	96	distill: + pseudo_llm_plus_v2strict.csv (Qwen3-14B v2 LLM etiketleri, 518K), seed 42
ce_bgeattrllm43/	BAAI/bge-reranker-v2-m3 (568M)	attr	96	aynı distill reçetesi seed 43

Güncel en iyi public dosya (0.91): submissions/ce_ps43llmwide_pos0.240.csv = ps43 (= l3bgeattrmmps+l3bgeattrmm43 50/50) tabanına Qwen3-14B rubrik-v2 LLM düzeltmesi (131K çelişki bandı + 186K geniş bant = 317K etiketli çift, data/processed/llm_labels_{contested,wide}.csv). Zincir: 0.88 (attr) → 0.89 (aile-ort) → 0.90 (LLM contested) → 0.91 (LLM geniş bant). Skor cache data/processed/l3ps43llmw_test_probs.npy. Modeller sadece taban ve distill ailesi; LLM düzeltmesi test-olasılığı seviyesinde (etiket CSV'leri GitHub LFS'te). Mirror model repo: bvrtuu/trendyol-eticaret-2026-models (aynı upload script, --repo-id ile).

Each folder is a standard Transformers AutoModelForSequenceClassification checkpoint with tokenizer files and a local meta.json.

Download

From the project repository:

HF_MODEL_REPO=efeyol11/trendyol-eticaret-2026-models \
uv run python scripts/download_models.py

The repo is public, so hf auth login is not required for download. Logging in can still help with rate limits.

To fetch every experiment folder on the Hub instead of the default set:

HF_MODEL_REPO=efeyol11/trendyol-eticaret-2026-models \
uv run python scripts/download_models.py --models all

To replace existing local model folders:

HF_MODEL_REPO=efeyol11/trendyol-eticaret-2026-models \
uv run python scripts/download_models.py --force

Usage

CE_OUT=models_store/ce CE_TAG=l3 uv run python scripts/predict_l3_cross_encoder.py
CE_OUT=models_store/ce_electra CE_TAG=l3el uv run python scripts/predict_l3_cross_encoder.py
CE_OUT=models_store/ce_rich CE_TAG=l3rich uv run python scripts/predict_l3_cross_encoder.py
CE_OUT=models_store/ce_hardft_faiss10 CE_TAG=l3hard uv run python scripts/predict_l3_cross_encoder.py
CE_OUT=models_store/ce_hardft_deep CE_TAG=l3deep uv run python scripts/predict_l3_cross_encoder.py
CE_OUT=models_store/ce_eldeep CE_TAG=l3eldeep uv run python scripts/predict_l3_cross_encoder.py
CE_OUT=models_store/ce_xlmr CE_TAG=l3xlmr uv run python scripts/predict_l3_cross_encoder.py

To upload or resync the public repository with a write-capable Hugging Face token:

hf auth login
HF_MODEL_REPO=efeyol11/trendyol-eticaret-2026-models \
uv run python scripts/upload_models_to_hf.py --public

Competition data and submission caches are shared through the private GitHub repository via Git LFS. Model checkpoints are shared here on Hugging Face.