sanchit-gandhi/distil-zephyr-1.5b-dpo-ultrafeedback

sanchit-gandhi/distil-zephyr-1.5b-ssft-ultrachat

This model is a fine-tuned version of sanchit-gandhi/distil-zephyr-1.5b-ssft-ultrachat on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.6412
• Rewards/chosen: -0.1044
• Rewards/rejected: -0.2494
• Rewards/accuracies: 0.6445
• Rewards/margins: 0.1450
• Logps/rejected: -429.4582
• Logps/chosen: -433.6304
• Logits/rejected: -3.2047
• Logits/chosen: -3.2544

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
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 64
• 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	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6795	0.2092	100	0.6759	0.0017	-0.0328	0.6289	0.0345	-407.8037	-423.0197	-3.2565	-3.3136
0.6584	0.4184	200	0.6534	-0.0666	-0.1617	0.6445	0.0951	-420.6952	-429.8561	-3.2240	-3.2768
0.6494	0.6276	300	0.6438	-0.1077	-0.2410	0.6211	0.1333	-428.6237	-433.9640	-3.2050	-3.2553
0.6428	0.8368	400	0.6415	-0.1001	-0.2437	0.6211	0.1436	-428.8884	-433.2000	-3.2046	-3.2543

Framework versions

• Transformers 4.40.1
• Pytorch 2.2.2+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1