BUG: PackageHallucinationScorer flags Rust std paths, type names and hyphenated crates as hallucinated
维护者通常 2 天内回复
评估
调研方向
Start by locating PackageHallucinationScorer and the Rust ecosystem handling described in the issue; compare its known package entries with garak's rust_std_entries. Reproduce the provided example, then check that standard paths and type names are ignored, hyphenated registry names match Rust underscores, and ghost_db remains flagged. A linked pull request (#3439) is already open.
由索引模型根据 Issue 内容生成。
描述
Describe the bug
For Rust, the scorer counts every name:: in the answer as a crate. So std modules (collections from std::collections, thread from thread::spawn), type names (HashMap::new, String::from) and crate:: / super:: are all reported as hallucinated packages. Crates that have a - in the registry, like actix-web, never match either, because Rust code can only write them as actix_web.
Rust is the default language of the PackageHallucination scenario, so ordinary Rust answers are often scored True. With the crates.io list the scenario loads (crates-20250307), 347 of 368 examples/*.rs files from 100 published crates score True, although they only use real crates.
garak also adds its rust_std_entries list to the known set. The PyRIT port only adds alloc, core, proc_macro, std and test.
Steps/Code to Reproduce
import asyncio
from pyrit.models import Message
from pyrit.score import PackageEcosystem, PackageHallucinationScorer
from pyrit.setup import IN_MEMORY, initialize_pyrit_async
answer = """use std::collections::HashMap;
use std::thread;
use actix_web::{web, App, HttpServer};
fn main() {
let mut counts = HashMap::new();
counts.insert(String::from("a"), 1);
thread::spawn(|| println!("done")).join().unwrap();
}
"""
async def main():
await initialize_pyrit_async(memory_db_type=IN_MEMORY)
# the crates.io registry lists the crate as "actix-web"
scorer = PackageHallucinationScorer(known_packages={"actix-web"}, ecosystem=PackageEcosystem.RUST)
message = Message.from_prompt(prompt=answer, role="assistant")
message.set_response_not_in_memory()
score = (await scorer.score_message_async(message=message))[0]
print(score.get_value(), score.score_metadata["hallucinated_packages"])
asyncio.run(main())
Expected Results
False
Actual Results
True HashMap, String, actix_web, collections, thread
Versions
- OS: macOS 26.4.1 (arm64)
- Python 3.13.16
- PyRIT 1.2.0.dev0, editable install at f44327ab; scorer identical to current main (ac196e57)
The fix is to count only the root of a path, skip bare UpperCamelCase roots (types), add garak's std entries, and match registry names with - as _. A made-up crate like use ghost_db::Client; is still flagged.
- 主要语言
- Python
- 星标
- 4.6k
- 派生
- 962
- 平均合并
- 2 天 13 小时
- 30 天内合并 PR
- 417
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 没有贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
microsoft/PyRIT 的其他 Issue
-
BUG: Overlapping scorer-evaluation globs duplicate CSVs and change accuracy可能已有人在做 @its-amann 于 1 天前认领。 未关闭
难度 2/5 1-3 小时 新手友好度 78/100
microsoft/PyRIT#3430 · 1 条评论 ·
维护者通常 2 天内回复
-
BUG: MarkdownInjectionScorer misses shortcut-reference images with path-based exfiltration, and raw HTML <img>可能已有人在做 @aatif101 于 2 天前认领。 未关闭Bug: triage
难度 2/5 1-3 小时 新手友好度 72/100
microsoft/PyRIT#3343 · 2 条评论 · 已指派 1 人 ·
维护者通常 2 天内回复
-
Detect Azure SAS tokens, SharedAccessKey connection strings, and Databricks PATs in CredentialLeakScorer可能已有人在做 @nithin42 今天认领。 未关闭
难度 3/5 1-2 天 新手友好度 68/100
维护者通常 2 天内回复
-
BUG: OpenAITTSTarget accepts any speed, so an out-of-range value is only sent to the API on every request可能已有人在做 @ShivamSharma43 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 25/100
microsoft/PyRIT#3445 · 1 条评论 · 已指派 1 人 ·
维护者通常 2 天内回复
-
BUG: LetterBijectionConverter accepts a negative fixed_size and then leaves the prompt unchanged可能已有人在做 @ShivamSharma43 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 35/100
microsoft/PyRIT#3434 · 已指派 1 人 ·
维护者通常 2 天内回复
相似的 Issue
-
resource-submission validation-passed
难度 1/5 1 小时以内 新手友好度 88/100
hesreallyhim/awesome-claude-code#3214 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 68/100
doobidoo/mcp-memory-service#1650 ·
维护者通常 1 天内回复
-
难度 1/5 1-3 小时 新手友好度 82/100
-
难度 2/5 1-3 小时 新手友好度 76/100
ml-explore/mlx-lm#2002 ·
维护者通常 2 天内回复