Modelpublic

tiffuhknee/qwen3-1.7b-newsroom-verifier

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes10downloads

Model Card

Cited Newsroom Verifier — Qwen3-1.7B (QLoRA adapter)

A LoRA adapter that turns Qwen/Qwen3-1.7B into a cited, evidence-bounded newsroom verifier: given a passage + retrieved sources, it flags each claim/quote as supported / unsupported / misleading and only says supported when it can quote the verbatim backing sentence from a real source — it never vouches for a claim it can't cite. (AP Style is handled deterministically in the app, not the model.)

Results (held-out golden set)

Model	spec_pass	verdict_acc	fabricated↓
This model (1.7B, tuned)	73.7–84%	74.9%	2.9–5.6%
GPT-OSS-120B (zero-shot)	30.3%	46.3%	6.1%
Llama-3.3-70B (zero-shot)	14.3%	25.7%	2.4%
Base Qwen3-1.7B (zero-shot)	9.1%	19.4%	61.4%

Beats a prompted 120B by ~2.4×. Base vs tuned significance: McNemar p < 0.0001. Thesis: behavior from data, not scale.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = "Qwen/Qwen3-1.7B"
tok = AutoTokenizer.from_pretrained(base)
m = AutoModelForCausalLM.from_pretrained(base)
m = PeftModel.from_pretrained(m, "tiffuhknee/qwen3-1.7b-newsroom-verifier")