Zeus v22
claim verification against text
1 min readupdated
Zeus v22 is a fact-checking model that takes a claim and several text passages, then determines whether they support, contradict, or don't cover it. It outputs an exact supporting quote along with a structured JSON verdict. Built on mistralai/Mistral-7B-Instruct-v0.2 with a LoRA (a lightweight adapter for fine-tuning a base model), it was trained on over 36,000 examples. It is designed for developers building information retrieval systems, fact-checking tools, and automated text reliability pipelines.
Key features
- Accurate claim classification (supported, contradicted, not_covered)
- Exact quote extraction from the source text as evidence
- Grounded strictly in the provided passages with no hallucinations about external facts
- Output formatted in strict JSON
What you can do
- Integrate it into fact-checking pipelines and LLM output verification workflows
- Analyze documents and cross-check claims against original sources
- Build claim verification services on top of knowledge bases
Example
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "Heliusnetwork/zeus")
tok = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")Read more
DetailsMetrics: 73.3% on VitaminC dev, 92.0% on Commons eval v3
Training: LoRA rank 8 on q_proj and v_proj, 60 minutes on an A100-40GB
License: Apache 2.0 (same as the base model)










Comments