hZzy/qwen2.5-0.5b-expo-DPO-EXPERIMENT-1W-5e6

qwen2.5-0.5b-expo-DPO-EXPERIMENT-1W

This model is a fine-tuned version of hZzy/qwen2.5-0.5b-sft-news-IFT on the hZzy/train_pairwise dataset. It achieves the following results on the evaluation set:

• Loss: 12497.9082
• Logps: -80.8773
• Logits: -1.3459
• Objective: 12906.1104
• Dpo Loss: 12906.1104
• Regularize: 12906.1104
• Ranking Simple: 0.5258
• Ranking Idealized: 0.5093
• Ranking Idealized Expo: 0.5093

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: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 6
• gradient_accumulation_steps: 12
• total_train_batch_size: 288
• total_eval_batch_size: 24
• 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	Logps	Logits	Objective	Dpo Loss	Regularize	Ranking Simple	Ranking Idealized	Ranking Idealized Expo
10470.1625	0.2834	50	11322.9775	-84.9886	-1.4586	11815.3389	11815.3389	11815.3389	0.5165	0.5093	0.5093
8558.2836	0.5668	100	12877.2754	-82.3385	-1.3720	13248.5166	13248.5166	13248.5166	0.5176	0.5093	0.5093
7521.95	0.8503	150	12664.2070	-80.5751	-1.3556	12926.7227	12926.7227	12926.7227	0.5227	0.5093	0.5093

Framework versions

• Transformers 4.42.0
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1