qwen2.5-0.5b-expo-DPO-EXPERIMENT-0.1-5e6

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: 0.7556
• Logps: -120.0429
• Logits: -1.9737
• Objective: 0.7840
• Dpo Loss: 0.7840
• Regularize: 0.7840
• Ranking Simple: 0.5403
• Ranking Idealized: 0.5888
• Ranking Idealized Expo: 0.5103

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
0.5807	0.2834	50	0.6822	-101.8611	-1.8709	0.7088	0.7088	0.7088	0.5196	0.5888	0.5103
0.4908	0.5668	100	0.6802	-105.0768	-1.8507	0.6854	0.6854	0.6854	0.5300	0.5888	0.5103
0.4191	0.8503	150	0.6960	-108.5704	-2.1205	0.7127	0.7127	0.7127	0.5403	0.5888	0.5103
0.2287	1.1337	200	0.7276	-115.4432	-2.0764	0.7403	0.7403	0.7403	0.5362	0.5888	0.5103
0.2329	1.4171	250	0.7454	-118.2405	-2.0640	0.7706	0.7706	0.7706	0.5351	0.5888	0.5103
0.2036	1.7005	300	0.7574	-120.7682	-1.9746	0.7851	0.7851	0.7851	0.5434	0.5888	0.5103
0.2102	1.9839	350	0.7556	-120.0429	-1.9737	0.7840	0.7840	0.7840	0.5403	0.5888	0.5103

Framework versions

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