Five different instruction-tuned models (which I'm sure are intuitively obvious from the name) merged using the methodology described in Resolving Interference When Merging Models.

In theory this should retain more of the capabilites of the constituent models than a straight linear merge would. In my testing, it feels quite capable.

Base model used for the merge: TheBloke/Llama-2-13B-fp16

Models merged in:

Works quite well with Alpaca-style prompts:

### Instruction:

...

### Response:

The script I used to perform the merge is available here.

The command that produced this model:

Detailed results can be found here