LoRA adapter for bert-base-uncased, fine-tuned on transcripts from the Huth Lab fMRI story-listening dataset for the Stat 214 (Spring 2026) final project at UC Berkeley.

The adapter is used to extract context-aware word embeddings that are then fed into a per-voxel ridge regression to predict whole-brain BOLD signal from spoken-story stimuli.

Configuration

Hyperparameter	Value
Base model	`bert-base-uncased`
LoRA rank `r`	4
LoRA alpha	8
LoRA dropout	0.1
Target modules	`query`, `value`
Training objective	Masked Language Modeling (MLM, 15%)
Training stories	86 (Huth Lab podcast transcripts)
MLM max sequence length	256
Epochs	3
Optimizer	AdamW, lr=2e-4
Batch size	16
Final MLM training loss	—

Encoding-model performance

After extracting per-word embeddings from this adapter (using ±10 word context windows + Lanczos downsampling + 4 TR delays) and fitting per-voxel ridge regression on Subjects 2 and 3:

Subject	Mean CC	Top 5% CC	Top 1% CC	Top-1 voxel
Subject 2	0.0643	0.2143	0.2906	0.4736
Subject 3	0.0660	0.2176	0.3043	0.5159

(See full project repository for ridge weights, evaluation code, and SHAP / LIME word-importance analyses.)

Loading the adapter

python

from transformers import BertForMaskedLM, BertTokenizerFast
from peft import PeftModel

tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
base = BertForMaskedLM.from_pretrained("bert-base-uncased")
model = PeftModel.from_pretrained(base, "RheaTinghe/stat214-lab3-bert-lora-r4-maxlen256")
model.eval()

# Extract per-word embeddings via ±10 word context windows
# (see scripts/run_bert_pretrained.py in the project repo for the
# complete extraction pipeline)

Citation

bibtex

@misc{stat214lab3,
  author = {Galloro, Drew and Wang, Ruihang and Khothsombath, Benjamin and Zhang, Rhea},
  title  = {Stat 214 Lab 3: BERT-LoRA encoding model for fMRI},
  year   = {2026},
  note   = {UC Berkeley Spring 2026},
}