Angshul/SpliNet-Embedder

SpliNet Embedder

Sentence-embedding version of SpliNet.

The complete encoder was fine-tuned end-to-end on 134,378 positive QQP duplicate pairs using symmetric contrastive learning.

Model

Property	Value
Encoder parameters	82,270,464
Embedding dimension	768
Maximum input length	512 tokens
Pooling	Masked mean
Output	L2-normalized
Fine-tuning pairs	134,378
Epochs	1
Optimization steps	8,958

Full evaluation

Benchmark	Metric	Score
STSBenchmark	Pearson	0.5257
STSBenchmark	Spearman	0.5174
SciFact	nDCG@10	0.1928
SciFact	MAP@10	0.1686
SciFact	MRR@10	0.1766
SciFact	Recall@10	0.2614
SciFact	Recall@100	0.4413
Twenty Newsgroups	Mean V-measure	0.1532
Twenty Newsgroups	Mean homogeneity	0.1510
Twenty Newsgroups	Mean completeness	0.1555

Twenty Newsgroups evaluation used 59,545 texts across all ten subsets.

Files

• splinet_embedder.ts — complete deployable TorchScript encoder
• splinet_embedder_full_qqp.pt — original trained checkpoint
• spiece.model — tokenizer
• api.py — Python/FastAPI interface
• model_info.json — metadata
• evaluation/ — benchmark results
• training/ — final training metadata

Usage

Install requirements.txt, then:

from api import encode, similarity
z = encode(['first sentence', 'second sentence'])
print(z.shape)

To serve:

python api.py

Endpoints: /health, /model-info, /embed, /similarity.

Base model: https://huggingface.co/Angshul/SpliNet

Source: https://github.com/AngshulMajumdar/SpliNet