qwen2.5-0.5b-expo-DPO-ES-100

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: 226.3468
• Logps: -80.2667
• Logits: -0.6269
• Objective: 213.3031
• Dpo Loss: 213.3031
• Regularize: 213.3031
• Ranking Simple: 0.5399
• Ranking Idealized: 0.5212
• Ranking Idealized Expo: 0.5212
• Wo Beta: 6.6215

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: 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: 5

Training results

Training Loss	Epoch	Step	Validation Loss	Logps	Logits	Objective	Dpo Loss	Regularize	Ranking Simple	Ranking Idealized	Ranking Idealized Expo	Wo Beta
17.3723	0.1417	50	32.1125	-90.9520	-1.4391	31.4854	31.4854	31.4854	0.5264	0.5212	0.5212	7.6851
60.4454	0.2834	100	70.9968	-86.6000	-1.4386	70.7719	70.7719	70.7719	0.5305	0.5212	0.5212	7.5289
100.2237	0.4251	150	129.8928	-85.7303	-1.2892	126.8845	126.8845	126.8845	0.5321	0.5212	0.5212	7.4641
120.8284	0.5668	200	164.0152	-75.5542	-1.3195	159.5013	159.5013	159.5013	0.5357	0.5212	0.5212	7.1836
134.8217	0.7085	250	195.7212	-79.3891	-1.2058	190.8510	190.8510	190.8510	0.5285	0.5212	0.5212	7.2711
119.0273	0.8503	300	192.5231	-84.2971	-0.9945	188.0580	188.0580	188.0580	0.5357	0.5212	0.5212	6.9382
114.0792	0.9920	350	205.7797	-82.1125	-1.0045	192.3920	192.3920	192.3920	0.5409	0.5212	0.5212	6.9235
72.4145	1.1337	400	212.6613	-82.8156	-0.7120	204.8122	204.8122	204.8122	0.5409	0.5212	0.5212	7.0485
76.9668	1.2754	450	210.2291	-82.4190	-0.7807	203.0261	203.0261	203.0261	0.5383	0.5212	0.5212	6.9244
77.9261	1.4171	500	211.3156	-81.3728	-0.7438	202.1569	202.1569	202.1569	0.5362	0.5212	0.5212	6.8863
70.5755	1.5588	550	212.6468	-82.3296	-0.6838	200.1410	200.1410	200.1410	0.5430	0.5212	0.5212	6.7241
69.6026	1.7005	600	212.0254	-80.7129	-0.5569	196.9669	196.9669	196.9669	0.5419	0.5212	0.5212	6.6975
69.7829	1.8422	650	222.2766	-79.4968	-0.7062	209.6782	209.6782	209.6782	0.5404	0.5212	0.5212	6.6541
62.7864	1.9839	700	226.3468	-80.2667	-0.6269	213.3031	213.3031	213.3031	0.5399	0.5212	0.5212	6.6215
37.3326	2.1256	750	219.7785	-80.5665	-0.7007	208.8723	208.8723	208.8723	0.5440	0.5212	0.5212	6.7265
33.2099	2.2674	800	221.8786	-81.8901	-0.5673	207.6881	207.6881	207.6881	0.5450	0.5212	0.5212	6.6717
33.915	2.4091	850	217.6955	-81.9134	-0.5178	205.0515	205.0515	205.0515	0.5424	0.5212	0.5212	6.7249
35.3572	2.5508	900	224.5402	-81.5880	-0.4729	214.5052	214.5052	214.5052	0.5435	0.5212	0.5212	6.8278
31.032	2.6925	950	225.2907	-80.3480	-0.5542	216.5803	216.5803	216.5803	0.5419	0.5212	0.5212	6.8429

Framework versions

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