zephyr-7b-sft-full-orpo

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.4701
• Rewards/chosen: -0.0364
• Rewards/rejected: -0.0499
• Rewards/accuracies: 0.6587
• Rewards/margins: 0.0135
• Logps/rejected: -0.9978
• Logps/chosen: -0.7282
• Logits/rejected: -2.9263
• Logits/chosen: -2.9434
• Nll Loss: 0.4357
• Log Odds Ratio: -0.6093
• Log Odds Chosen: 0.4456

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: 7e-06
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: inverse_sqrt
• lr_scheduler_warmup_steps: 100
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Nll Loss	Log Odds Ratio	Log Odds Chosen
0.5226	0.1049	100	0.5280	-0.0386	-0.0472	0.6329	0.0086	-0.9448	-0.7728	-2.7583	-2.7860	0.4953	-0.6326	0.2873
0.5074	0.2098	200	0.5134	-0.0381	-0.0478	0.6409	0.0098	-0.9566	-0.7612	-2.6736	-2.7002	0.4774	-0.6357	0.3190
0.5265	0.3146	300	0.5012	-0.0379	-0.0479	0.6329	0.0099	-0.9572	-0.7588	-2.7317	-2.7594	0.4653	-0.6374	0.3278
0.5194	0.4195	400	0.4912	-0.0371	-0.0478	0.6429	0.0107	-0.9559	-0.7417	-2.6640	-2.6974	0.4560	-0.6284	0.3607
0.5008	0.5244	500	0.4847	-0.0373	-0.0489	0.6508	0.0117	-0.9786	-0.7455	-2.5957	-2.6294	0.4499	-0.6209	0.3873
0.4725	0.6293	600	0.4794	-0.0362	-0.0470	0.6349	0.0107	-0.9394	-0.7248	-2.6147	-2.6477	0.4435	-0.6320	0.3567
0.4875	0.7341	700	0.4767	-0.0368	-0.0498	0.6409	0.0129	-0.9955	-0.7365	-2.6910	-2.7213	0.4416	-0.6158	0.4180
0.4796	0.8390	800	0.4740	-0.0371	-0.0508	0.6508	0.0137	-1.0162	-0.7416	-2.7913	-2.8114	0.4396	-0.6169	0.4363
0.4851	0.9439	900	0.4714	-0.0357	-0.0466	0.6528	0.0109	-0.9324	-0.7143	-2.9543	-2.9692	0.4361	-0.6245	0.3669

Framework versions

• Transformers 4.41.0.dev0
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1