KBhandari11/centroid-adapter-reasonir

Centroid Adapter — reasonir

Lightweight BottleneckResidualAdapter trained on top of reasonir embeddings to produce representation-invariant table embeddings.

Architecture

z = e + α · Up( Dropout( GELU( Down( LN(e) ) ) ) )

Hyperparameter	Value
Embedding dim d	4096
Bottleneck rank r	512
Residual scale α	0.01
Use bias	True

Trained on: NQ, WTQ, WIKISQL

Usage

import torch
from huggingface_hub import hf_hub_download
import json

# --- option A: use the from_pretrained helper in this repo ---
# (copy BottleneckResidualAdapter + from_pretrained from push_to_hub.py)
adapter = BottleneckResidualAdapter.from_pretrained("KBhandari11/centroid-adapter-reasonir")
e = torch.randn(1, 4096)   # your backbone embedding
z = adapter(e)                    # representation-invariant embedding

# --- option B: hf_hub_download one-liner ---
from safetensors.torch import load_file
weights_path = hf_hub_download("KBhandari11/centroid-adapter-reasonir", "model.safetensors")
cfg_path     = hf_hub_download("KBhandari11/centroid-adapter-reasonir", "config.json")
with open(cfg_path) as f:
    cfg = json.load(f)
adapter = BottleneckResidualAdapter(**cfg)
adapter.load_state_dict(load_file(weights_path))
adapter.eval()

Research Paper

Improving Robustness of Tabular Retrieval via Representational Stability