rb512/resonant-frank

Resonant Frank — BGE-large + Tone Head

Frank is a retrieval + tone-aware embedding model. It pairs a frozen BGE-large-en-v1.5 backbone (competitive retrieval) with a trained tone head that separates register (formal vs casual) as an independent signal.

Architecture

• Backbone: BAAI/bge-large-en-v1.5 (335M params, frozen)
• Tone head v1: MLP on CLS token (657K params)
• Tone head v2: Multi-layer attention pooling from layers [16,18,20,22,24] (659K params) ← recommended

Retrieval uses BGE's native CLS path. Tone head is a separate read-only projection. No interference.

Results

Retrieval (BEIR, NDCG@10)

Task	Score
NFCorpus	0.604
SciFact	0.750
ArguAna	0.460
FiQA	0.568
Average	0.595

Tone Separation (length-controlled)

Version	Gap (matched)	Probe Acc
v1 (CLS)	0.361	98.5%
v2 (multi-layer)	0.493	96.9%

Task-Level Reranking (SciFact, target=casual)

Alpha	Register Match	nDCG@10
0.0 (BGE only)	8.8%	0.904
0.2 (v2 recommended)	97.1%	0.899

Usage

from transformers import AutoModel, AutoTokenizer
import torch, torch.nn as nn, torch.nn.functional as F
from huggingface_hub import hf_hub_download

# Load backbone
tok = AutoTokenizer.from_pretrained("BAAI/bge-large-en-v1.5")
backbone = AutoModel.from_pretrained("BAAI/bge-large-en-v1.5").eval()

# Load tone head (v2 recommended)
ckpt = torch.load(hf_hub_download("rb512/resonant-frank", "frank_v2_tone.pt"), map_location="cpu")
# See ToneBranchV2 class definition in the repo for architecture

License

• BGE-large-en-v1.5: MIT
• Tone head weights: MIT