--- license: apache-2.0 base_model: mistralai/Mistral-7B-Instruct-v0.2 library_name: peft pipeline_tag: text-generation tags: - fact-checking - claim-verification - lora - olo language: - en --- # Zeus v22: open claim verification Zeus reads a claim and a few passages and says whether the passages **support** it, **contradict** it or **do not cover** it. It quotes the exact words that carry the evidence. It is the checker behind [Olo](https://helius-olo.io), which tests every sentence of an answer against sources from the open [Olo Commons](https://heliusnetwork.app/olo). It is built in the open: the data pipeline, training code and every measured score are in [Hashim57/olo-commons](https://github.com/Hashim57/olo-commons) under `zeus/`. ## Measured scores Measured on 2026-11-16 for this exact version by `zeus/eval.py`. Nothing here is estimated. The held-out sets come first because the frozen first split has been at or near 100 percent for weeks and no longer shows progress. | Set | Label accuracy | |---|---| | VitaminC dev (300 human-written claims, held out) | 73.3% | | Commons eval v3 (frozen, hard patterns, n=200) | 92.0% | | Commons eval v2 (frozen, harder split, n=200) | 100.0% | | Commons eval (frozen first split, n=200) | 99.5% | Evidence hit on the frozen split (quoted span found in the right passage): 99.5%. JSON valid: 100.0%. ## Training - Base model: `mistralai/Mistral-7B-Instruct-v0.2` (Apache 2.0). - Method: LoRA rank 8 on q_proj, v_proj, learning rate 0.0002, batch 16, 1034 optimiser steps in 60.0 minutes on one A100-40GB. - Training examples: 36,255. - Trained by the public GitHub workflow `zeus-train.yml` on a cloud GPU; the full log of every version is `zeus/STATUS.md`. ## Data and licences Three parts, described in [`zeus/DATA_CARD.md`](https://github.com/Hashim57/olo-commons/blob/main/zeus/DATA_CARD.md): examples made by rule from Olo Commons atoms (each atom keeps its own source licence), hand-written gold examples checked against the atoms they quote, and human-written claim and evidence pairs from VitaminC (Schuster, Fisch and Barzilay, NAACL 2021, CC BY-SA 3.0). Text from Wikipedia and VitaminC is CC BY-SA: attribute Wikipedia and the VitaminC authors. The adapter is released under Apache 2.0, like its base model. ## Use ```python 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") ``` Send one user turn with the system prompt from `zeus/cloud/recipe.py` followed by: ``` Claim: Passages: [p1] () [p2] () ``` The answer is one JSON object: `{"label": "supported" | "contradicted" | "not_covered", "evidence": [{"passage": "", "quote": ""}], "rationale": ""}`. ## Limits Most training contradictions are made by rule (one changed relation, year, number or name), so subtle contradictions are harder for it than clear ones; the weakest pattern in each version is reported in the weekly learning report. It checks claims against the passages it is given and knows nothing beyond them. It is not a substitute for reading the source: every Olo answer carries its passages so a person can. Card generated by `zeus/cloud/publish_hf.py` on 2026-11-17 from `zeus/out/zeus-v22/`.