license: apache-2.0
This is a linear model merge of:
60% https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2
40% https://huggingface.co/stabilityai/japanese-stablelm-instruct-gamma-7b
I recommend following the Mistral chat template and prompting in English.
Evaluation
Tested on correct en-jp translation identification on the first 10k rows of https://huggingface.co/datasets/Verah/tatoeba_dedupe_en-jp_2024-March-01
Desired behaviour is to not accept any translation when we deliberaly test incorrect pairings from the dataset, and to not reject any translation when shown only correctly paired examples.
| Model | False Admissions | False Rejections |
|---|---|---|
| Mistral Instruct | 41 | 600 |
| (This Model) | 13 | 1839 |
| JP Stable LM Gamma | 9679 | 138 |
| Hermes2DPO | 20 | 598 |
I made the test harder by concatenating 3 paired sentences together, in the false admissions case 1 out of those 3 was incorrectly paired.
| Model | False Admissions | False* Rejections |
|---|---|---|
| (This Model) | 89 | 5508 |
| Hermes2DPO | 537 | 1458 |
This model also wanted to reject many "correct" translations, however 3 unrelated sentences back to back isn't a very correct thing to be doing, either.