- This is a merged model. So you have to unckeck the "[X] Contains merge/moerge" checkbox above the list.
| Metric | Value |
|---|---|
| Avg. | 81.28 |
- I was inspired by this Sakana project
[ ModelSelector ] > [ Finetunner ] > [ Merger ] > [ Evaluator ]
sequenceDiagram
participant Finetunner
participant ModelSelector
participant Merger
participant Evaluator
Finetunner->>ModelSelector: Tunned models
loop ScoreChecker
ModelSelector->>Merger: Highest two
Merger->>Evaluator: Merged one
Evaluator->>ModelSelector: If the score is not enough
end
Required two models with the same architecture.
- Original model > DPO + SFT > Gapped model
- SLERP Original one and Gapped one > Merged
- Evaluate > Merged
- Fine-tune a specific evaluation part of the model if you need to increase the score for that part
- SLERP Merged one and Original One > Second Merged
- Evaluate Second Merged
@ Continue until the average evaluation score is higher than the original one
- QWEN2