lr-experiment1-7B

The lr-experiment model series is a research project I'm conducting that I will be using to determine the best learning rate to use while fine-tuning Mistral. This model uses a learning rate of 2e-5 with a cosine scheduler and no warmup steps.

I used Locutusque/Hercules-2.0-Mistral-7B as a base model, and further fine-tuned it on CollectiveCognition/chats-data-2023-09-22 using QLoRA for 3 epochs. I will be keeping track of evaluation results, and will comparing it to upcoming models.

Evals

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
agieval_nous	N/A	none	None	acc	0.3645	±	0.0093
		none	None	acc_norm	0.3468	±	0.0092
- agieval_aqua_rat	1	none	None	acc	0.2283	±	0.0264
		none	None	acc_norm	0.2283	±	0.0264
- agieval_logiqa_en	1	none	None	acc	0.2965	±	0.0179
		none	None	acc_norm	0.3303	±	0.0184
- agieval_lsat_ar	1	none	None	acc	0.2217	±	0.0275
		none	None	acc_norm	0.1783	±	0.0253
- agieval_lsat_lr	1	none	None	acc	0.4039	±	0.0217
		none	None	acc_norm	0.3686	±	0.0214
- agieval_lsat_rc	1	none	None	acc	0.4870	±	0.0305
		none	None	acc_norm	0.4424	±	0.0303
- agieval_sat_en	1	none	None	acc	0.6408	±	0.0335
		none	None	acc_norm	0.5971	±	0.0343
- agieval_sat_en_without_passage	1	none	None	acc	0.3932	±	0.0341
		none	None	acc_norm	0.3835	±	0.0340
- agieval_sat_math	1	none	None	acc	0.3455	±	0.0321
		none	None	acc_norm	0.2727	±	0.0301

Groups	Version	Filter	n-shot	Metric	Value		Stderr
agieval_nous	N/A	none	None	acc	0.3645	±	0.0093
		none	None	acc_norm	0.3468	±	0.0092