Heliusnetwork commited on
Commit
25638ae
·
verified ·
1 Parent(s): 0f3ec00

Zeus v22: adapter and model card from measured files

Browse files
Files changed (3) hide show
  1. README.md +71 -0
  2. adapter_config.json +30 -0
  3. adapter_model.safetensors +3 -0
README.md ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: mistralai/Mistral-7B-Instruct-v0.2
4
+ library_name: peft
5
+ pipeline_tag: text-generation
6
+ tags:
7
+ - fact-checking
8
+ - claim-verification
9
+ - lora
10
+ - olo
11
+ language:
12
+ - en
13
+ ---
14
+
15
+ # Zeus v22: open claim verification
16
+
17
+ 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/`.
18
+
19
+ ## Measured scores
20
+
21
+ 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.
22
+
23
+ | Set | Label accuracy |
24
+ |---|---|
25
+ | VitaminC dev (300 human-written claims, held out) | 73.3% |
26
+ | Commons eval v3 (frozen, hard patterns, n=200) | 92.0% |
27
+ | Commons eval v2 (frozen, harder split, n=200) | 100.0% |
28
+ | Commons eval (frozen first split, n=200) | 99.5% |
29
+
30
+ Evidence hit on the frozen split (quoted span found in the right passage): 99.5%. JSON valid: 100.0%.
31
+
32
+ ## Training
33
+
34
+ - Base model: `mistralai/Mistral-7B-Instruct-v0.2` (Apache 2.0).
35
+ - 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.
36
+ - Training examples: 36,255.
37
+ - Trained by the public GitHub workflow `zeus-train.yml` on a cloud GPU; the full log of every version is `zeus/STATUS.md`.
38
+
39
+ ## Data and licences
40
+
41
+ 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.
42
+
43
+ ## Use
44
+
45
+ ```python
46
+ from transformers import AutoModelForCausalLM, AutoTokenizer
47
+ from peft import PeftModel
48
+
49
+ base = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", torch_dtype="auto", device_map="auto")
50
+ model = PeftModel.from_pretrained(base, "Heliusnetwork/zeus")
51
+ tok = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
52
+ ```
53
+
54
+ Send one user turn with the system prompt from `zeus/cloud/recipe.py` followed by:
55
+
56
+ ```
57
+ Claim: <claim>
58
+
59
+ Passages:
60
+ [p1] (<source>) <text>
61
+
62
+ [p2] (<source>) <text>
63
+ ```
64
+
65
+ The answer is one JSON object: `{"label": "supported" | "contradicted" | "not_covered", "evidence": [{"passage": "<id>", "quote": "<exact words>"}], "rationale": "<one sentence>"}`.
66
+
67
+ ## Limits
68
+
69
+ 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.
70
+
71
+ Card generated by `zeus/cloud/publish_hf.py` on 2026-11-17 from `zeus/out/zeus-v22/`.
adapter_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alpha_pattern": {},
3
+ "auto_mapping": null,
4
+ "base_model_name_or_path": "mistralai/Mistral-7B-Instruct-v0.2",
5
+ "bias": "none",
6
+ "fan_in_fan_out": false,
7
+ "inference_mode": true,
8
+ "init_lora_weights": true,
9
+ "layer_replication": null,
10
+ "layers_pattern": null,
11
+ "layers_to_transform": null,
12
+ "loftq_config": {},
13
+ "lora_alpha": 16,
14
+ "lora_dropout": 0.05,
15
+ "megatron_config": null,
16
+ "megatron_core": "megatron.core",
17
+ "modules_to_save": null,
18
+ "peft_type": "LORA",
19
+ "r": 8,
20
+ "rank_pattern": {},
21
+ "revision": null,
22
+ "target_modules": [
23
+ "v_proj",
24
+ "q_proj"
25
+ ],
26
+ "task_type": "CAUSAL_LM",
27
+ "use_dora": false,
28
+ "use_rslora": false,
29
+ "model_type": "mistral"
30
+ }
adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc53605ca3fa53a252f1dd2d384efd5d219588ee6d0a4da244d70349ef58574d
3
+ size 13648432