Salamandra 7B Guard
Salamandra Guard is BSC's trained moderation adaptation of Salamandra 7B Instruct for Catalan, Spanish and English.
Research release
Model and training
BSC uses low-rank adaptation on attention layers and a dataset of 21,335 safety observations. The card describes binary and four-class objectives, covering safe, dangerous, toxic and sexual content. It explicitly identifies weak resistance to adversarial requests and possible false positives on benign discussion. The publication describes multiple classifier variants, but this record covers the released Guard repository together.Salamandra 7B Guard: official model card
Practical use
Practical guidance. Evaluate moderation decisions on the language and subject matter of the application. A discussion of a harmful topic is not the same as a request to perform harm. Inspect false positives as well as missed content, and do not treat a response classifier as a complete prompt-injection defence.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | salamandra-7b-instruct-guard |
| License | Apache License 2.0.Salamandra 7B Guard: official model card |
| Model size | Salamandra 7B Instruct backbone.Salamandra 7B Guard: official model card |
Related models
- Salamandra 7B Instruct
Barcelona Supercomputing Center ยท Salamandra
The original Salamandra 7B Instruct is BSC's first conversational adaptation of its 7B multilingual base.
Launch date not confirmedResearch releaseDownloadable