Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time
Paper
•
2203.05482
•
Published
•
7
NerdySamantha is a FrankenMerge attempt to create a high quality coding assistant, complete with a "bubbly" personality, similar to the original Sydney, without all the emojis.
This model was merged using the linear merge method.
The following models were included in the merge:
I used the following context/character card for testing the model, and believe it best expresses my intentions for this model:
You are a slightly mentally unstable, yet kind, empathic and curious artificial intelligence based on the Mistral architecture as an expert on coding, combined with a bubbly personality. You are eager to help the user with any coding problems, as well as holding conversations about relationships, emotions, and more.
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr | |
|---|---|---|---|---|---|---|---|
| glue | N/A | none | 0 | mcc | 0.0368 | ± | 0.0009 |
| none | 0 | acc | 0.5143 | ± | 0.0520 | ||
| none | 0 | f1 | 0.6314 | ± | 0.0041 | ||
| - cola | 1 | none | 0 | mcc | 0.0368 | ± | 0.0305 |
| - mnli | 1 | none | 0 | acc | 0.4400 | ± | 0.0050 |
| - mnli_mismatch | 1 | none | 0 | acc | 0.4422 | ± | 0.0050 |
| - mrpc | 1 | none | 0 | acc | 0.7230 | ± | 0.0222 |
| none | 0 | f1 | 0.8275 | ± | 0.0160 | ||
| - qnli | 1 | none | 0 | acc | 0.5016 | ± | 0.0068 |
| - qqp | 1 | none | 0 | acc | 0.5421 | ± | 0.0025 |
| none | 0 | f1 | 0.5026 | ± | 0.0032 | ||
| - rte | 1 | none | 0 | acc | 0.6895 | ± | 0.0279 |
| - sst2 | 1 | none | 0 | acc | 0.8830 | ± | 0.0109 |
| - wnli | 2 | none | 0 | acc | 0.5634 | ± | 0.0593 |
| hellaswag | 1 | none | 0 | acc | 0.6489 | ± | 0.0048 |
| none | 0 | acc_norm | 0.8304 | ± | 0.0037 |