README.md

---
library_name: transformers
tags:
- merge
- mergekit
license: cc-by-4.0
language:
- da
- sv
- 'no'
- nb
- nn
base_model:
- mistralai/Mistral-7B-v0.1
- danish-foundation-models/munin-7b-alpha
- norallm/normistral-7b-warm
- timpal0l/Mistral-7B-v0.1-flashback-v2
---

# ScandiMerge

This is a DARE-TIES merge of the following models, all based on `mistralai/Mistral-7B-v0.1`:

1. `danish-foundation-models/munin-7b-alpha`, continued pretraining on Danish data;
2. `norallm/normistral-7b-warm`, continued pretraining on Norwegian data;
3. `timpal0l/Mistral-7B-v0.1-flashback-v2`, continued pretraining on Swedish data.



## Model Details

- **Merged by:** Dan Saattrup Nielsen
- **Model type:** Decoder model, based on Mistral-7B-v0.1
- **Language(s) (NLP):** Danish, Swedish and Norwegian
- **License:** CC-BY-4.0
- **Merge configuration:**
  ```python
  dict(
      models=[
          dict(
              model="danish-foundation-models/munin-7b-alpha",
              parameters=dict(
                  density=0.9,
                  weight=1.0,
              ),
          ),
          dict(
              model="norallm/normistral-7b-warm",
              parameters=dict(
                  density=0.9,
                  weight=1.0,
              ),
          ),
          dict(
              model="timpal0l/Mistral-7B-v0.1-flashback-v2",
              parameters=dict(
                  density=0.9,
                  weight=1.0,
              ),
          ),
      ],
      merge_method="dare_ties",
      random_seed=4242,
      base_model="mistralai/Mistral-7B-v0.1",
      parameters=dict(
          normalize=True,
          int8_mask=True,
      ),
      dtype="float16",
  )
  ```