Purdy0228/ConvMemory-ZH-DualSpace-GTE

ConvMemory Chinese Dual-Space GTE Student

What This Checkpoint Is

This is the representative checkpoint for the Chinese ConvMemory dual-space reranker.

The online inference structure is:

query/memory text -> base GTE embedding + triplet-GTE embedding -> normalized dual-space vector -> ConvMemory window encoder -> CE-lite reranker

It is not an online cross-encoder. The Jina reranker was used only as an offline teacher during training.

Intended Use

Use this checkpoint as a plug-and-play Chinese retrieval/reranking module through the convmemory package:

from convmemory import ChineseConvMemory

model = ChineseConvMemory.from_pretrained("Purdy0228/ConvMemory-ZH-DualSpace-GTE")
ranked = model.rerank("用户最近喜欢什么音乐？", memories, top_k=10)

This is a Chinese retrieval/reranking checkpoint, not the validity-context module.

Assets

• Lightweight student weights: student.pt
• Config and provenance: config.json, MANIFEST.json
• Base encoder dependency: Alibaba-NLP/gte-multilingual-base
• Bundled tuned encoder: triplet_encoder/

The Hub repository bundles the tuned SentenceTransformer under triplet_encoder/, so users only need the repo id plus the base encoder dependency. student.pt stores the lightweight ConvMemory student; the tuned encoder is packaged beside it for plug-and-play loading.

Five-Seed Method-Level Result

selection	R@10	Hit@10	MRR
best by R@10	0.7871 +/- 0.0043	0.8400 +/- 0.0051	0.6145 +/- 0.0026
best by MRR	0.7855 +/- 0.0060	0.8388 +/- 0.0062	0.6153 +/- 0.0025
fixed rw=0	0.7867 +/- 0.0045	0.8397 +/- 0.0051	0.6149 +/- 0.0028

Released Representative Checkpoint

The released checkpoint is a representative seed-31 model selected to be close to the five-seed mean under fixed rw=0.

Package-level checkpoint-load evaluation, no training:

method	R@10	Hit@10	MRR
celite_rerank_top500_rw0	0.787233	0.839506	0.615516
celite_rerank_top500_rw0.025	0.788169	0.840629	0.613753

Checksums

• student.pt SHA256: e634950ae53fb4a8ad48e78fba86dbc26f52ac1b04fe1bef81f9887bd3ea6ec0