Aider Polyglot
Harbor nativeAider Polyglot: 6种语言diff编辑
Verified71%Coding
- Samples
- 225
- Released
- 2024
- Scores
- 61
Cross-Lingual Unlearning Tensor
CROSS-LINGUAL UNLEARNING TENSOR
A multilingual unlearning benchmark spanning 174 language-script pairs and 25 paraphrase types to study cross-lingual knowledge erasure.
How to run
ltb · cross-lingual-unlearning-tensor
$ npm i -g lettokenburn-cli$ ltb bench info cross-lingual-unlearning-tensor
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
8 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.4-nano | 84.9 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-4o | 87.1 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-5-mini | 87.2 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-5.1 | 88.0 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-4.1 | 88.3 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| DeepSeek-V3 | 88.9 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-5 | 94.4 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| GPT-5.2-chat | 94.8 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
Content & Construction: The dataset spans 174 language–script pairs across high, medium, and low-resource tiers. It includes forget samples re-expressed through 25 atomic paraphrase types in seven meta-categories, with translations into 40 languages. Models are fine-tuned on synthetic tasks (TOFU) and real-world news documents from the LORELEI corpus.
Scoring & Key Findings: The benchmark evaluates unlearning via oracle normalized residual access across different query-answer routes (native vs cross-lingual). It introduces COVER, a method for selecting source languages to maximize coverage of held-out languages. Results show that naive selection fails compared to COVER, which reduces mean held-out residual access by 7.8–27.3%.