nidhipandya/NewsScope-lora

NewsScope LoRA Adapter

This repository contains a LoRA adapter fine-tuned for schema-grounded claim extraction from news articles.

It produces structured JSON outputs with:

• domain
• headline
• key_points
• whos_involved
• how_it_unfolded
• claims (2-3 verifiable claims with evidence)

Key Result (Human Evaluation)

• NewsScope: 89.4% accuracy
• GPT-4o-mini baseline: 93.7%
• Reported difference is not statistically significant (p=0.07)

Important: LLaMA License

You must accept the Meta LLaMA license for the base model on Hugging Face:
meta-llama/Meta-Llama-3.1-8B-Instruct

Then either:

• run huggingface-cli login, or
• set HF_TOKEN in your environment.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3.1-8B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, "nidhipandya/NewsScope-lora")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")

Training Details

• Base model: meta-llama/Meta-Llama-3.1-8B-Instruct
• LoRA rank: 16
• Training set size: 315 articles (URLs + annotations; article text not publicly redistributed)
• Notes: Training reproduction requires fetching article text from URLs due to copyright.

Links

• Code: https://github.com/nidhip1611/NewsScope
• Benchmark: GitHub Releases (benchmark.zip)
• Paper: arXiv (TBD)

Citation

@article{pandyaNewsscope,
  title={NewsScope: Schema-Grounded Cross-Domain News Claim Extraction with Open Models},
  author={Pandya, Nidhi},
  journal={arXiv preprint arXiv:TBD},
  year={TBD}
}