flax-lora-distilbert-base-uncased

LoRA Adapter for distilbert-base-uncased

This adapter was trained using Flax NNX for the task of sentiment classification on the IMDB dataset.

Configuration

• LoRA rank: 16
• LoRA alpha: 32
• Target modules: ['q_lin', 'v_lin']

Use

# Loading and using the adapter
import orbax.checkpoint as ocp
import pickle

# Method 1: Orbax (recommended for larger models)
checkpointer = ocp.StandardCheckpointer()
lora_params = checkpointer.restore("lora_checkpoint")

# Method 2: Pickle (easier for smaller models)
with open("lora_params.pkl", "rb") as f:
    lora_params = pickle.load(f)

# Reconstruction of LoRA layers
from flax import nnx
lora_layers = None
for layer_name, params in lora_params.items():
    # Create a new LoRA layer with the correct parameters
    # (implementation depends on the specific architecture)
    pass