File size: 3,521 Bytes
25638ae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
---
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: <claim>

Passages:
[p1] (<source>) <text>

[p2] (<source>) <text>
```

The answer is one JSON object: `{"label": "supported" | "contradicted" | "not_covered", "evidence": [{"passage": "<id>", "quote": "<exact words>"}], "rationale": "<one sentence>"}`.

## 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/`.