qwen2.5-0.5b-expo-L2EXPO-W2-noES-0.1

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

• Loss: 0.0378
• Logps: -86.2505
• Logits: -1.3514
• Objective: 0.0371
• Dpo Loss: 0.6773
• Regularize: 0.4086
• Ranking Simple: 0.5367
• Ranking Idealized: 0.6025
• Ranking Idealized Expo: 0.5233
• Wo Beta: 16.0267

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: 1e-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 3
• gradient_accumulation_steps: 12
• total_train_batch_size: 144
• total_eval_batch_size: 12
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Logps	Logits	Objective	Dpo Loss	Regularize	Ranking Simple	Ranking Idealized	Ranking Idealized Expo	Wo Beta
0.0369	0.1417	50	0.0390	-90.2782	-1.4491	0.0391	0.6884	0.4122	0.5269	0.6025	0.5233	16.4652
0.0352	0.2834	100	0.0391	-91.7039	-1.5424	0.0391	0.6823	0.4173	0.5285	0.6025	0.5233	16.4180
0.0307	0.4251	150	0.0381	-89.9212	-1.4936	0.0378	0.6776	0.4083	0.5305	0.6025	0.5233	16.7819
0.0297	0.5668	200	0.0383	-89.7068	-1.4461	0.0379	0.6807	0.4114	0.5311	0.6025	0.5233	16.3721
0.0308	0.7085	250	0.0393	-84.8614	-1.4727	0.0387	0.6793	0.4269	0.5373	0.6025	0.5233	16.4728
0.0319	0.8503	300	0.0393	-86.9588	-1.3799	0.0385	0.6849	0.4346	0.5362	0.6025	0.5233	16.5196
0.0332	0.9920	350	0.0385	-85.9522	-1.4162	0.0371	0.6765	0.4138	0.5367	0.6025	0.5233	16.1268
0.0316	1.1337	400	0.0381	-85.4246	-1.3429	0.0375	0.6780	0.4129	0.5367	0.6025	0.5233	16.0028
0.0231	1.2754	450	0.0383	-86.0492	-1.2800	0.0377	0.6798	0.4181	0.5383	0.6025	0.5233	16.0724
0.0276	1.4171	500	0.0385	-85.1352	-1.3222	0.0376	0.6785	0.4148	0.5362	0.6025	0.5233	15.9100
0.0264	1.5588	550	0.0381	-85.1091	-1.3170	0.0375	0.6773	0.4145	0.5367	0.6025	0.5233	16.0355
0.0216	1.7005	600	0.0379	-84.8271	-1.3649	0.0373	0.6764	0.4106	0.5378	0.6025	0.5233	15.9760
0.0221	1.8422	650	0.0381	-86.0721	-1.3792	0.0371	0.6776	0.4128	0.5367	0.6025	0.5233	15.9955
0.0208	1.9839	700	0.0383	-85.6073	-1.3779	0.0375	0.6793	0.4142	0.5357	0.6025	0.5233	15.9613
0.0149	2.1256	750	0.0379	-86.0851	-1.3526	0.0372	0.6783	0.4108	0.5342	0.6025	0.5233	16.0025
0.0157	2.2674	800	0.0378	-86.2671	-1.3483	0.0370	0.6781	0.4103	0.5362	0.6025	0.5233	16.0414
0.0173	2.4091	850	0.0377	-86.2224	-1.3534	0.0370	0.6770	0.4081	0.5383	0.6025	0.5233	16.0221
0.018	2.5508	900	0.0378	-86.2056	-1.3507	0.0371	0.6773	0.4091	0.5373	0.6025	0.5233	15.9948
0.0133	2.6925	950	0.0378	-86.1726	-1.3498	0.0371	0.6775	0.4090	0.5367	0.6025	0.5233	16.0199
0.0126	2.8342	1000	0.0378	-86.2452	-1.3515	0.0371	0.6773	0.4086	0.5373	0.6025	0.5233	16.0244
0.0146	2.9759	1050	0.0378	-86.2505	-1.3514	0.0371	0.6773	0.4086	0.5367	0.6025	0.5233	16.0267

Framework versions

• Transformers 4.42.0
• Pytorch 2.3.0+cu121
• Datasets 3.2.0
• Tokenizers 0.19.1