DanielJeongsooLee/adaptive-msc-v7

Adaptive MSC V7 Model

This repository contains the trained checkpoint for the Adaptive Multi-Step Context (MSC) model (Version 7). MSC is designed to improve RAG efficiency by dynamically deciding when to stop reading chunks and which chunks to prune.

How to Use

1. Requirements

Ensure you have the adaptive-msc library installed.

2. Download Checkpoint

You can download the checkpoint directly using huggingface_hub:

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="DanielJeongsooLee/adaptive-msc-v7",
    filename="v7_msc_e3.pth"
)
print(f"Downloaded to: {checkpoint_path}")

3. Load in Code

import torch
from adaptive_msc.models.msc_model import AdaptiveMSCModelV7

model = AdaptiveMSCModelV7(model_name="microsoft/deberta-v3-small")
checkpoint = torch.load(checkpoint_path, map_location="cpu")
model.load_state_dict(checkpoint)
model.eval()