Mistral2_1000_STEPS_05beta_CDPOSFT

This model is a fine-tuned version of tsavage68/mistralit2_1000_STEPS_5e7_SFT on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 1.5804
• Rewards/chosen: 1.1615
• Rewards/rejected: 0.9028
• Rewards/accuracies: 0.4286
• Rewards/margins: 0.2587
• Logps/rejected: -75.7158
• Logps/chosen: -73.1790
• Logits/rejected: -1.8951
• Logits/chosen: -1.8951

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-05
• train_batch_size: 4
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 1000

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
2.0435	0.0977	50	1.6546	-0.2111	-0.0540	0.3868	-0.1572	-77.6294	-75.9242	-1.6195	-1.6195
3.1098	0.1953	100	1.7215	-0.1670	-0.5209	0.4286	0.3539	-78.5632	-75.8359	-0.7416	-0.7416
1.8949	0.2930	150	1.6841	2.1671	1.9786	0.4154	0.1886	-73.5644	-71.1677	-1.7619	-1.7619
1.4406	0.3906	200	1.6936	2.3177	2.1054	0.4264	0.2124	-73.3108	-70.8665	-2.2879	-2.2879
1.5623	0.4883	250	1.5911	0.8418	0.4811	0.4396	0.3607	-76.5593	-73.8184	-1.5834	-1.5834
1.8884	0.5859	300	1.5747	1.4552	1.2105	0.4418	0.2447	-75.1005	-72.5916	-1.6640	-1.6640
1.4373	0.6836	350	1.5569	1.3020	1.0909	0.4198	0.2111	-75.3397	-72.8979	-1.9137	-1.9136
1.4732	0.7812	400	1.5216	1.0023	0.6676	0.4571	0.3347	-76.1863	-73.4973	-1.9794	-1.9794
1.9109	0.8789	450	1.5502	1.3520	0.9986	0.4505	0.3534	-75.5243	-72.7979	-1.8076	-1.8076
1.4744	0.9766	500	1.5531	1.3605	1.1014	0.4264	0.2591	-75.3186	-72.7809	-1.9385	-1.9385
1.2615	1.0742	550	1.6623	0.6530	0.4114	0.4242	0.2415	-76.6986	-74.1960	-2.3949	-2.3949
1.8019	1.1719	600	1.6240	0.8707	0.6200	0.4308	0.2507	-76.2815	-73.7606	-1.6149	-1.6149
1.2202	1.2695	650	1.5993	1.1246	0.9014	0.4330	0.2233	-75.7188	-73.2527	-1.8964	-1.8964
1.0924	1.3672	700	1.5922	1.3888	1.1674	0.4242	0.2214	-75.1866	-72.7243	-1.8455	-1.8455
0.8059	1.4648	750	1.6004	1.1205	0.8834	0.4396	0.2371	-75.7547	-73.2610	-1.9415	-1.9415
0.9489	1.5625	800	1.5917	1.2725	1.0232	0.4264	0.2493	-75.4751	-72.9570	-1.9293	-1.9293
1.2564	1.6602	850	1.5797	1.1856	0.9286	0.4264	0.2570	-75.6643	-73.1308	-1.8894	-1.8894
1.2613	1.7578	900	1.5806	1.1682	0.9110	0.4308	0.2572	-75.6995	-73.1655	-1.8963	-1.8963
1.1197	1.8555	950	1.5804	1.1615	0.9030	0.4286	0.2585	-75.7156	-73.1791	-1.8955	-1.8955
0.7665	1.9531	1000	1.5804	1.1615	0.9028	0.4286	0.2587	-75.7158	-73.1790	-1.8951	-1.8951

Framework versions

• Transformers 4.40.1
• Pytorch 2.0.0+cu117
• Datasets 2.19.0
• Tokenizers 0.19.1