Necent/efficientrag-labeler-mdeberta-v3-base

EfficientRAG Labeler (mdeberta-v3-base)

Labeler component of EfficientRAG — dual-headed DeBERTa model for multi-hop retrieval.

What it does

Given a query and a retrieved chunk, the Labeler:

1. Sequence classification: Is this chunk relevant (CONTINUE) or irrelevant (TERMINATE)?
2. Token classification: Which tokens in the chunk are useful for answering?

Architecture

• Base: microsoft/mdeberta-v3-base (86M params, multilingual)
• Custom dual head: DebertaForSequenceTokenClassification
  • Token head: binary per-token (useful/useless)
  • Sequence head: binary per-chunk (CONTINUE/TERMINATE)

Training

Data	30,818 samples (HotpotQA EN + Dragon-derec RU)
Epochs	2
Batch size	4
LR	5e-6
Max length	384
Hardware	Apple M3 Pro, ~3.4 hours

Usage

Results on DRAGON benchmark

Metric	Baseline	EfficientRAG	Delta
MRR (multi-hop)	0.736	0.798	+0.062
MRR (overall)	0.783	0.822	+0.040
Precision	0.187	0.582	+0.395

Related

• Training data: Necent/efficientrag-labeler-training-data
• Filter model: Necent/efficientrag-filter-mdeberta-v3-base
• Paper: EfficientRAG (arXiv:2408.04259)