Modelpublic

tsavage68/400STEPS_5e7rate_03beta_DPO_Meditron7B_zeroshot

sourceHugging Facellama2updated 3y agoView on Hugging Face

1likes18downloads

Model Card

400STEPS5e7rate03betaDPOMeditron7B

This model is a fine-tuned version of epfl-llm/meditron-7b on an unknown dataset. It achieves the following results on the evaluation set:

—Loss: 0.6439
—Rewards/chosen: -0.0166
—Rewards/rejected: -0.1472
—Rewards/accuracies: 0.5714
—Rewards/margins: 0.1306
—Logps/rejected: -28.2845
—Logps/chosen: -26.5367
—Logits/rejected: -0.6342
—Logits/chosen: -0.6341

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

—learning_rate: 5e-07
—trainbatchsize: 4
—evalbatchsize: 1
—seed: 42
—gradientaccumulationsteps: 2
—totaltrainbatch_size: 8
—optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
—lrschedulertype: cosine
—lrschedulerwarmup_steps: 100
—training_steps: 400

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6896	0.1	50	0.6916	0.0067	0.0033	0.4637	0.0034	-27.7828	-26.4590	-0.6113	-0.6111
0.6783	0.2	100	0.6771	-0.0693	-0.1071	0.5319	0.0378	-28.1508	-26.7125	-0.6173	-0.6171
0.6697	0.29	150	0.6571	-0.0107	-0.1001	0.5626	0.0893	-28.1273	-26.5172	-0.6171	-0.6170
0.6463	0.39	200	0.6496	0.0037	-0.1067	0.5692	0.1104	-28.1493	-26.4691	-0.6288	-0.6286
0.6124	0.49	250	0.6449	-0.0073	-0.1329	0.5648	0.1257	-28.2368	-26.5056	-0.6318	-0.6317
0.641	0.59	300	0.6440	-0.0156	-0.1460	0.5758	0.1304	-28.2803	-26.5333	-0.6340	-0.6339
0.643	0.68	350	0.6430	-0.0150	-0.1479	0.5780	0.1328	-28.2866	-26.5315	-0.6343	-0.6341
0.6632	0.78	400	0.6439	-0.0166	-0.1472	0.5714	0.1306	-28.2845	-26.5367	-0.6342	-0.6341

Framework versions

—Transformers 4.37.2
—Pytorch 2.0.0+cu117
—Datasets 2.17.0
—Tokenizers 0.15.1