qwen2.5-0.5b-expo-DPO-EXPERIMENT-1-5e6

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: 1.5850
• Logps: -81.4617
• Logits: -1.1346
• Objective: 1.6058
• Dpo Loss: 1.6058
• Regularize: 1.6058
• Ranking Simple: 0.5217
• Ranking Idealized: 0.5145
• 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: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Logps	Logits	Objective	Dpo Loss	Regularize	Ranking Simple	Ranking Idealized	Ranking Idealized Expo
1.0653	0.2834	50	1.1911	-88.4918	-1.4769	1.2503	1.2503	1.2503	0.5186	0.5145	0.5093
1.0433	0.5668	100	1.6201	-82.0376	-1.4625	1.6224	1.6224	1.6224	0.5227	0.5145	0.5093
0.8546	0.8503	150	1.6978	-82.1610	-1.2570	1.7001	1.7001	1.7001	0.5258	0.5145	0.5093
0.3915	1.1337	200	1.7222	-81.7287	-1.2002	1.7483	1.7483	1.7483	0.5155	0.5145	0.5093
0.361	1.4171	250	1.6479	-81.3939	-1.1553	1.6762	1.6762	1.6762	0.5269	0.5145	0.5093
0.2219	1.7005	300	1.5928	-81.4689	-1.1283	1.6152	1.6152	1.6152	0.5248	0.5145	0.5093
0.2782	1.9839	350	1.5850	-81.4617	-1.1346	1.6058	1.6058	1.6058	0.5217	0.5145	0.5093

Framework versions

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