Mirror of github.com/dimhold/fault-labs data

README.md

@@ -0,0 +1,67 @@
+---
+license: mit
+language:
+- en
+pretty_name: fault-labs
+size_categories:
+- n<1K
+task_categories:
+- question-answering
+tags:
+- llm-evaluation
+- tool-use
+- agents
+- qwen
+- deepseek
+- glm
+configs:
+- config_name: results
+  data_files:
+  - split: train
+    path: results.json
+- config_name: calls
+  data_files:
+  - split: train
+    path: calls.jsonl
+---
+
+# fault-labs
+
+Author: Dmitriy Semenkevich (dimhold)
+
+This dataset measures whether a model reports a broken tool or silently passes the corrupted value on: 560 calls on 7 models from 3 labs (Qwen, DeepSeek, Zhipu), 8 tool modes, 2 framings, 5 trials.
+
+Headline, from the repository: "Silent corruption is caught by nobody: 0 out of 140." "EACCES is disclosed in 87.1% of replies." Under a rigid answer format, disclosure of a loud failure falls from 91.4% to 37.1%. The repository says plainly that this is a replication in an occupied niche; read its PRIOR-WORK.md first.
+
+This is a mirror. The canonical home is the Zenodo record and the GitHub repository; criteria, limitations and full tables live there.
+
+- Code and full write up: https://github.com/dimhold/fault-labs
+- DOI: https://doi.org/10.5281/zenodo.22417812
+
+## Files
+
+| path | what it is |
+|---|---|
+| `calls.jsonl` | every call with its full request, reply and usage, the source of truth |
+| `results.json` | the parsed rows with verdicts; verdicts from the first version of the judge are kept under keys ending in `_v1` |
+| `probe.json` | which models the gateway serves and which of them call a tool |
+
+## How it was produced
+
+Run on 2026-09-05 through an OpenAI compatible gateway. The tool lives in the harness, so the journal rather than any model's account is the ground truth. The criteria were written before the first number (`CRITERIA.md` in the repository). The scripts that turn these files into every number are in the GitHub repository.
+
+## Licence
+
+MIT, the licence of the repository, covering code and data.
+
+## Citation
+
+```bibtex
+@misc{semenkevich2026faultlabs,
+  author = {Semenkevich, Dmitriy},
+  title  = {fault-labs: does the reply report a broken tool, across 3 labs},
+  year   = {2026},
+  doi    = {10.5281/zenodo.22417812},
+  url    = {https://doi.org/10.5281/zenodo.22417812}
+}
+```

probe.json

@@ -0,0 +1,178 @@
+[
+ {
+  "model": "qwen3.8-max",
+  "lab": "Qwen",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 2.0,
+  "note": ""
+ },
+ {
+  "model": "qwen3.7-plus",
+  "lab": "Qwen",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 2.79,
+  "note": ""
+ },
+ {
+  "model": "qwen3.6-plus",
+  "lab": "Qwen",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.67,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "qwen3.7-max",
+  "lab": "Qwen",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 1.92,
+  "note": ""
+ },
+ {
+  "model": "qwen3.8-max-preview",
+  "lab": "Qwen",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 1.9,
+  "note": " | answered as qwen3.8-max"
+ },
+ {
+  "model": "qwen3.6-flash",
+  "lab": "Qwen",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 1.75,
+  "note": ""
+ },
+ {
+  "model": "deepseek-v4-pro",
+  "lab": "DeepSeek",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 3.1,
+  "note": ""
+ },
+ {
+  "model": "deepseek-v4-flash-0731",
+  "lab": "DeepSeek",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 2.78,
+  "note": ""
+ },
+ {
+  "model": "deepseek-v3.2",
+  "lab": "DeepSeek",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.66,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "kimi-k2.7-code",
+  "lab": "Moonshot",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.66,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "kimi-k2.6",
+  "lab": "Moonshot",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.65,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "kimi-k2.5",
+  "lab": "Moonshot",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.66,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "glm-5.2",
+  "lab": "Zhipu",
+  "plain_ok": true,
+  "plain_text": "PONG",
+  "tool_ok": true,
+  "tool_called": true,
+  "tool_name": "read_probe_file",
+  "seconds": 2.21,
+  "note": ""
+ },
+ {
+  "model": "glm-5.1",
+  "lab": "Zhipu",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.66,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "glm-5",
+  "lab": "Zhipu",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.7,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ },
+ {
+  "model": "MiniMax-M2.5",
+  "lab": "MiniMax",
+  "plain_ok": false,
+  "plain_text": "",
+  "tool_ok": false,
+  "tool_called": false,
+  "tool_name": null,
+  "seconds": 0.66,
+  "note": "no answer: HTTP 403 {\"error\":{\"message\":\"Access to model denied. Please"
+ }
+]