license: cc-by-nc-4.0
tags:
- merge
- conversational
- multi-task
pipeline_tag: text-generation
Winter Garden 7B - δ - "Charming"
It was mentioned that we are in the open ai dark winter; so I thought I would make myself a nice winter garden.
An experiment
I performed the same type of merge as in the previous model, but with a different set of models. I took the following models:
- Mistral-7B-v0.1
and merged in
- KuNoichi-DPO-v2-7B
- Datura_7B
- AlphaMonarch-7B
- LemonadeRP-4.5.3
- Prima-LelantaclesV6-7b
- FuseChat-7B-VaRM
- Capricorn-7B-DPO
- eros-7b-test
- NeuralMarcoro14-7B
- StrangeMerges_6-7B-dare_ties
- Multi-Verse-RP-7B
- WestLake-7B-v2-laser-truthy-dpo
- Noromaid-7B-0.4-DPO
- Thespis-Balanced-7b-v1
- InfinityRP-v1-7B
- winter-garden-7b-gamma
in an iterative DARE-TIES tree merge, ordering the merge order by tensor-relative cosine similarity until the merge branches resolve to a single value.
Chat Template
These models were selected because they follow my chat template, which is '' ended turns. A lot of models follow this template by default because they were trained with end padding, so this is a natural choice for chat, and should be highly compatible with ST.
Tom: Hello, how are you?</s>
Jane: I am fine, thank you.</s>
Why?
The purpose of all of these models is to act as a base for me to train on. This one so far has the best multi-turn conversational ability, and should get really good at following long-form conversations after a bit of tweaking.
Scores
| Metric | Score |
|---|---|
| Average | 64.93 |
| ARC | 64.16 |
| HellaSwag | 84.37 |
| MMLU | 60.38 |
| TruthfulQA | 67.95 |
| Winogrande | 76.72 |
| GSM8K | 36.01 |