You can deploy this model and start earning money today!

Purdy0228/
ConvMemory-OPC-Student-BGE

Name: ConvMemory-OPC-Student-BGE
Availability: InStock

sentence-transformers

Views

No views yet

ConvMemory OPC Student BGE

This checkpoint is an OPC-style Chinese ConvMemory student reranker. It keeps the ConvMemory inference pattern while using a Chinese-specialized BGE encoder:

bge-base-zh embeddings -> ConvMemory window encoder -> CE-lite reranker head.

Usage

python

from convmemory import OPCConvMemory

model = OPCConvMemory.from_pretrained(
    "Purdy0228/ConvMemory-OPC-Student-BGE",
    device="cuda",  # or "cpu"
)

ranked = model.rerank(
    query="现在定价方案是什么？",
    memories=[
        {"id": "m1", "text": "定价方案最终是基础版 99 元，专业版 299 元。"},
        {"id": "m2", "text": "获客渠道第一批先做小红书。"},
        {"id": "m3", "text": "旧方案曾经考虑过 49 元，后来被否掉。"},
    ],
    top_k=3,
)

Architecture

Dense encoder: BAAI/bge-base-zh-v1.5
ConvMemory window encoder
CE-lite student reranker head
Offline teacher used for distillation only: BAAI/bge-reranker-v2-m3

The teacher is not used at inference time. Evaluation uses construction-gold memories from the synthetic OPC-style data generator, not teacher agreement.

Intended Use

Use this checkpoint for Chinese OPC-style memory retrieval when candidate pools are large enough that an online cross-encoder teacher is too slow.

For small candidate pools, a strong dense retriever such as BAAI/bge-base-zh-v1.5 may be sufficient. The phase-0 gate did not show a statistically robust improvement over BGE cosine retrieval, so this checkpoint should be treated as a ConvMemory student operating point rather than a universal replacement for the base retriever.

Released API Evaluation

Package path: OPCConvMemory.from_pretrained(...)

split	questions	R@1	R@5	R@10	Hit@1	Hit@10	MRR
phase-0 pass-only test	43	0.4516	0.9583	0.9913	0.7442	1.0000	0.8484

Warm-Memory Pool Scan

pool size	teacher MRR	student MRR	teacher ms/query	student ms/query
10	0.9302	0.9380	19.89	5.54
50	0.8702	0.8411	48.24	26.60
200	0.8302	0.8070	166.53	30.17
500	0.7760	0.7445	390.79	35.66
1000	0.6803	0.6599	779.67	126.18

Files

student.pt: ConvMemory window encoder and CE-lite scorer weights
config.json: model and architecture config
MANIFEST.json: provenance and checksums

Checksums

student.pt SHA256: e56a74d937446d05de7381307f9188cffb8794e07805ee9837838ed8503713eb
config.json SHA256: 24e838d0e86bd839f618e51890a4330d9d677bdda790d27bea53ed98614c9b77
training/evaluation data SHA256: c8f595ee8afd2aa2d1a2f6336ac172f297667e3ec3b69bd7f8e876b925a1319d