qwen2.5-0.5b-expo-L1EXPO-ES-10

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: 51.6794
• Logps: -80.6186
• Logits: -0.5140
• Objective: 51.5628
• Dpo Loss: 26.2794
• Regularize: 51.5628
• Ranking Simple: 0.5248
• Ranking Idealized: 0.5212
• Ranking Idealized Expo: 0.5212
• Wo Beta: 14.0744

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	Dpo Loss	Logits	Logps	Validation Loss	Objective	Ranking Idealized	Ranking Idealized Expo	Ranking Simple	Regularize	Wo Beta
4.2778	0.1417	50	2.8787	-1.4301	-91.7813	5.6511	5.5786	0.5212	0.5212	0.5243	5.5786	16.1070
17.3516	0.2834	100	7.9687	-1.3171	-86.6635	15.6835	15.7548	0.5212	0.5212	0.5280	15.7548	15.6261
28.6009	0.4251	150	15.0002	-1.1259	-81.4986	29.0753	28.9045	0.5212	0.5212	0.5243	28.9045	15.2369
35.0698	0.5668	200	21.3918	-0.8776	-82.1578	41.1263	40.4593	0.5212	0.5212	0.5124	40.4593	14.9112
37.7822	0.7085	250	21.9288	-0.6419	-83.0039	44.0746	43.3933	0.5212	0.5212	0.5280	43.3933	14.6204
35.2811	0.8503	300	21.4307	-0.5316	-83.8429	43.6626	43.4643	0.5212	0.5212	0.5321	43.4643	14.5447
33.8034	0.9920	350	23.3301	-0.5934	-84.0573	45.2649	45.3586	0.5212	0.5212	0.5238	45.3586	14.6023
30.8702	1.1337	400	23.8270	-0.6271	-82.2022	47.2698	47.2674	0.5212	0.5212	0.5248	47.2674	14.3367
29.5027	1.2754	450	25.1794	-0.5508	-82.7233	49.3412	49.4737	0.5212	0.5212	0.5202	49.4737	14.3433
27.7693	1.4171	500	24.6274	-0.5208	-83.1404	48.4138	48.5616	0.5212	0.5212	0.5181	48.5616	14.3259
26.3455	1.5588	550	24.8876	-0.5377	-81.6711	49.4754	49.7513	0.5212	0.5212	0.5264	49.7513	14.2335
25.3777	1.7005	600	24.6279	-0.5633	-81.3699	48.8078	49.2645	0.5212	0.5212	0.5238	49.2645	14.1972
24.4429	1.8422	650	25.3419	-0.4757	-81.6565	49.7105	49.8172	0.5212	0.5212	0.5192	49.8172	14.3368
22.5358	1.9839	700	26.2794	-0.5140	-80.6186	51.6794	51.5628	0.5212	0.5212	0.5248	51.5628	14.0744
20.6864	2.1256	750	25.7920	-0.4511	-83.9474	50.9028	51.1398	0.5212	0.5212	0.5274	51.1398	14.2847
19.5881	2.2674	800	26.2232	-0.4519	-84.1413	51.4440	51.8351	0.5212	0.5212	0.5274	51.8351	14.2120
18.5246	2.4091	850	26.5269	-0.5061	-82.9639	52.2825	52.2313	0.5212	0.5212	0.5285	52.2313	14.1205
17.4115	2.5508	900	26.5477	-0.5079	-83.9889	52.2686	52.2795	0.5212	0.5212	0.5290	52.2795	14.1975
16.2052	2.6925	950	26.6571	-0.4691	-83.1267	52.4042	52.3891	0.5212	0.5212	0.5238	52.3891	14.2985
15.0384	2.8389	1000	51.7636	-82.8277	-0.4551	51.6447	26.1645	51.6447	0.5264	0.5212	0.5212	14.2036
14.381	2.9806	1050	51.8214	-83.0540	-0.4122	51.9024	26.5043	51.9024	0.5248	0.5212	0.5212	14.1669
12.5437	3.1223	1100	51.6017	-83.8731	-0.4408	51.8998	26.1851	51.8998	0.5254	0.5212	0.5212	14.1769
11.3828	3.2641	1150	51.5869	-84.2104	-0.4506	51.7268	26.2023	51.7268	0.5259	0.5212	0.5212	14.1768
10.5152	3.4058	1200	51.5859	-84.1485	-0.4568	51.6626	26.3073	51.6626	0.5254	0.5212	0.5212	14.1450

Framework versions

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