TeeZee/GALAXY-XB-v.03
Experiment, can DUS be taken one or more steps further?
Technical notes:
- 12 layers removed from both models, 4 more than in original paper but its 1/4 of all layers(48) as per original paper.
- base version of upstage/SOLAR-10.7B-v1.0 used for merge
- no finetuning done yet, this is just a merge, first step in DUS paper
- next step, if evaluation proves that its at least as 'smart' as base model, should be finetuning to 'recover' after merge
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 63.37 |
| AI2 Reasoning Challenge (25-Shot) | 61.77 |
| HellaSwag (10-Shot) | 83.59 |
| MMLU (5-Shot) | 64.55 |
| TruthfulQA (0-shot) | 44.19 |
| Winogrande (5-shot) | 81.06 |
| GSM8k (5-shot) | 45.03 |
Results
- small quality loss can be observed comparing to base model, as described in the DUS paper
- this merge has best evaluation results, so it will be finetuned to 'recover' from the merge
- finetunig will be done on 5-10% of openorca dataset and full DPO datasets used by SOLAR
- v03 > v01 > v02 - based on average evaluation scores, removing 1/4 of total layers seems to be the correct way to scale DUS