zephyr-7b-dpo-oursuf6k-qlora-5e-6

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the generation/UF6k dataset. It achieves the following results on the evaluation set:

• Loss: 0.5972
• Rewards/chosen: -2.1562
• Rewards/rejected: -2.9617
• Rewards/accuracies: 0.6865
• Rewards/margins: 0.8055
• Rewards/margins Max: 2.4253
• Rewards/margins Min: -0.7592
• Rewards/margins Std: 1.4237
• Logps/rejected: -555.3531
• Logps/chosen: -500.8411
• Logits/rejected: -1.8009
• Logits/chosen: -1.8449

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-06
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 16
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Rewards/margins Max	Rewards/margins Min	Rewards/margins Std	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5821	0.15	100	0.6622	-0.1867	-0.2755	0.6151	0.0888	0.3953	-0.1795	0.2588	-286.7321	-303.8942	-2.6966	-2.7355
0.481	0.29	200	0.6257	-1.2575	-1.6473	0.6746	0.3898	1.3412	-0.5259	0.8283	-423.9109	-410.9716	-2.5402	-2.5661
0.4017	0.44	300	0.6112	-1.7680	-2.5016	0.6944	0.7336	2.3329	-0.8123	1.4011	-509.3477	-462.0217	-1.9880	-2.0224
0.3427	0.58	400	0.5955	-1.9140	-2.6859	0.7024	0.7719	2.2721	-0.7218	1.3401	-527.7765	-476.6219	-1.9447	-1.9863
0.3246	0.73	500	0.6026	-2.2815	-3.0194	0.6627	0.7379	2.2879	-0.7821	1.3716	-561.1234	-513.3748	-1.8444	-1.8864
0.2747	0.88	600	0.5973	-2.1734	-2.9762	0.6786	0.8029	2.4273	-0.7515	1.4233	-556.8073	-502.5607	-1.7934	-1.8380

Framework versions

• PEFT 0.7.1
• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2