Salamandra 7B Instruct 2606
Salamandra 7B Instruct 2606 updates BSC's conversational model with longer context and preliminary preference alignment.
Model and training
BSC describes improvements in system instructions, extended conversations and resistance to noisy inputs. It uses synthetic preference pairs for helpfulness and safety. The card's architecture table incorrectly lists 40B-sized values; the released configuration instead specifies 32 layers and a 4,096-wide hidden state. Its configured position limit is 163,840 tokens.Salamandra 7B Instruct 2606: official model cardSalamandra 7B Instruct 2606: released configuration
Practical use
Practical guidance. Evaluate the release with documents that contain conflicting instructions and with conversations that change topic. A longer configured window does not prove that all positions contribute equally to the answer. Compare short and long inputs, and check whether the model can cite the part of the supplied text that supports its response.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | salamandra-7b-instruct-2606 |
| Context window | 163,840 tokens Maximum position setting in the released configuration; effective long-document accuracy needs task-specific evaluation. Salamandra 7B Instruct 2606: released configuration |
| License | Apache License 2.0.Salamandra 7B Instruct 2606: official model cardSalamandra 7B Instruct 2606: released configuration |
| Model size | 7B model class; the card contains an inconsistent 40B architecture table.Salamandra 7B Instruct 2606: official model cardSalamandra 7B Instruct 2606: released configuration |
Related models
- Salamandra 7B FC 2607
Barcelona Supercomputing Center · Salamandra
Salamandra 7B FC 2607 adapts BSC's 2606 instruction model for function calling in a smaller model class.
Launch date not confirmedRestricted accessDownloadable The original Salamandra 7B Instruct is BSC's first conversational adaptation of its 7B multilingual base.
Launch date not confirmedResearch releaseDownloadable