mariklolik/ecup-2026-qilowoq-soft-human-final
E-CUP 2026 Qilowoq soft-to-human reranker
Exact continuation checkpoint behind ODS submission qilowoq-soft-global-final-12c515e5.zip.
Verified result
- ODS Public Macro PR-AUC:
0.47069991672790107 - ODS status:
Success - submission archive SHA-256:
12c515e5112dd526d227cfb62dc3764a27a67d0c0ad4cd841121ba236c1c5e9e - model weights SHA-256:
eae1e3af1734e4e44ea5d438656dd93126b06af258a36605de99928e7cbf263d
Architecture and training
Base checkpoint: qilowoq/bge-reranker-v2-m3-en-ru at revision 0ca0a6a717ffbe01e25c6e8e33bad7982d938930.
The model is a sequence-classification cross-encoder over a canonicalized pair of product texts. Product text contains category, title, and normalized attributes. Item IDs are used only to make pair orientation deterministic and never enter model text.
Training has two stages:
- one epoch at max length 128 on a target-blind, category-balanced 1,000,000-pair soft-label sample;
- one epoch at max length 384 on all 329,088 component-disjoint development human pairs.
Both stages use batch size 64, learning rate 4e-6, BF16 autocast, FP32 logits/loss, AdamW, global positive weighting, category weighting, gradient clipping, deterministic seed 42, and alternating pair-swap augmentation.
The checkpoint is saved as FP32 safetensors and reloaded before packaging. training-provenance.json, reranker-spec.json, and model-manifest.json bind every source and training artifact to SHA-256.
Reproduction code
- training entry point
- training and artifact contract
- pair serialization and provenance validation
- inference runner
- package builder
- C1 evidence using this checkpoint
The GitHub repository and this model repository are private team assets.