zephyr-7b-dpo-qlora-no-sft

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5486
• Rewards/chosen: -1.4557
• Rewards/rejected: -2.2032
• Rewards/accuracies: 0.7090
• Rewards/margins: 0.7475
• Logps/rejected: -484.1859
• Logps/chosen: -430.8606
• Logits/rejected: 0.8536
• Logits/chosen: 0.8124

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: 1
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6934	0.01	100	0.6930	0.0021	0.0018	0.5120	0.0003	-263.6873	-285.0847	-2.5761	-2.6081
0.6921	0.03	200	0.6923	0.0064	0.0047	0.5820	0.0017	-263.3970	-284.6488	-2.5766	-2.6089
0.6913	0.04	300	0.6910	0.0127	0.0083	0.6195	0.0044	-263.0383	-284.0253	-2.5774	-2.6105
0.6888	0.05	400	0.6894	0.0235	0.0157	0.6210	0.0077	-262.2991	-282.9474	-2.5778	-2.6114
0.6881	0.07	500	0.6866	0.0322	0.0186	0.6220	0.0136	-262.0058	-282.0685	-2.5648	-2.6011
0.6848	0.08	600	0.6829	0.0391	0.0173	0.6230	0.0218	-262.1442	-281.3836	-2.5621	-2.6006
0.6706	0.09	700	0.6776	0.0515	0.0169	0.6135	0.0346	-262.1758	-280.1425	-2.5437	-2.5861
0.6544	0.1	800	0.6650	-0.0843	-0.1603	0.6065	0.0760	-279.8956	-293.7216	-2.5208	-2.5676
0.668	0.12	900	0.6552	-0.1689	-0.2798	0.6170	0.1109	-291.8528	-302.1819	-2.5180	-2.5613
0.6285	0.13	1000	0.6457	-0.5287	-0.7121	0.6290	0.1834	-335.0806	-338.1635	-2.4563	-2.4939
0.6741	0.14	1100	0.6396	-0.7030	-0.9481	0.6305	0.2452	-358.6847	-355.5893	-2.2815	-2.3227
0.605	0.16	1200	0.6279	-0.7077	-0.9713	0.6375	0.2636	-360.9963	-356.0601	-2.2198	-2.2608
0.5844	0.17	1300	0.6228	-0.8502	-1.1414	0.6410	0.2912	-378.0121	-370.3147	-2.0337	-2.0743
0.6085	0.18	1400	0.6157	-0.6163	-0.8963	0.6565	0.2799	-353.4970	-346.9268	-1.9276	-1.9742
0.5887	0.2	1500	0.6093	-1.0534	-1.4085	0.6585	0.3551	-404.7234	-390.6338	-1.5130	-1.5476
0.5585	0.21	1600	0.6020	-0.8558	-1.2372	0.6645	0.3814	-387.5893	-370.8767	-1.4216	-1.4652
0.5417	0.22	1700	0.5937	-0.7787	-1.1648	0.6640	0.3860	-380.3489	-363.1672	-1.3190	-1.3614
0.5691	0.24	1800	0.5964	-1.0690	-1.5628	0.6705	0.4938	-420.1472	-392.1945	-0.7433	-0.7891
0.5869	0.25	1900	0.5931	-1.4234	-1.8618	0.6700	0.4384	-450.0478	-427.6318	-0.5757	-0.5963
0.6732	0.26	2000	0.5928	-0.7320	-1.1323	0.6765	0.4002	-377.0961	-358.4945	-0.8928	-0.9596
0.5453	0.27	2100	0.5812	-1.2215	-1.6723	0.6770	0.4508	-431.1005	-407.4461	-0.3057	-0.3325
0.5521	0.29	2200	0.5773	-0.9855	-1.4907	0.6775	0.5052	-412.9417	-383.8439	-0.0835	-0.1059
0.5352	0.3	2300	0.5821	-1.0780	-1.5279	0.6885	0.4500	-416.6599	-393.0880	-0.2117	-0.2432
0.4291	0.31	2400	0.5800	-1.3780	-1.9871	0.6785	0.6091	-462.5805	-423.0901	0.1802	0.1741
0.5324	0.33	2500	0.5709	-1.0291	-1.5875	0.6765	0.5584	-422.6171	-388.1980	0.0904	0.0751
0.5659	0.34	2600	0.5640	-1.2533	-1.8232	0.6985	0.5699	-446.1898	-410.6243	0.3281	0.3241
0.5041	0.35	2700	0.5737	-1.7469	-2.3921	0.6865	0.6452	-503.0828	-459.9810	0.5911	0.5924
0.5754	0.37	2800	0.5716	-1.6382	-2.2298	0.6885	0.5915	-486.8488	-449.1171	0.6424	0.6612
0.6073	0.38	2900	0.5731	-1.5512	-2.2130	0.6815	0.6618	-485.1724	-440.4115	0.7017	0.6979
0.6283	0.39	3000	0.5645	-1.3105	-1.9937	0.6860	0.6832	-463.2372	-416.3378	0.6221	0.5951
0.5199	0.41	3100	0.5585	-1.1618	-1.7386	0.6940	0.5768	-437.7283	-401.4741	0.4404	0.4092
0.5658	0.42	3200	0.5603	-1.1916	-1.7704	0.6960	0.5788	-440.9099	-404.4548	0.3075	0.2535
0.6214	0.43	3300	0.5605	-1.3366	-1.9673	0.6925	0.6307	-460.5986	-418.9480	0.6742	0.6564
0.581	0.44	3400	0.5563	-1.1359	-1.7683	0.6985	0.6324	-440.7018	-398.8812	0.5839	0.5449
0.5422	0.46	3500	0.5590	-1.0364	-1.6150	0.6915	0.5786	-425.3734	-388.9318	0.5735	0.5330
0.5626	0.47	3600	0.5602	-1.1120	-1.7501	0.6910	0.6381	-438.8792	-396.4902	0.7862	0.7520
0.627	0.48	3700	0.5579	-1.2845	-1.9488	0.6935	0.6644	-458.7537	-413.7391	0.8809	0.8576
0.5522	0.5	3800	0.5562	-1.3810	-2.0706	0.6975	0.6896	-470.9312	-423.3916	0.9118	0.8745
0.5734	0.51	3900	0.5557	-1.3964	-2.0908	0.6970	0.6943	-472.9462	-424.9361	0.7969	0.7417
0.612	0.52	4000	0.5548	-1.6249	-2.3232	0.7075	0.6982	-496.1850	-447.7854	0.8941	0.8718
0.5357	0.54	4100	0.5587	-1.1962	-1.8866	0.6995	0.6904	-452.5338	-404.9135	0.5836	0.5102
0.5648	0.55	4200	0.5570	-1.3147	-2.0461	0.6940	0.7314	-468.4804	-416.7626	0.7063	0.6440
0.5237	0.56	4300	0.5515	-1.5027	-2.2087	0.7030	0.7060	-484.7385	-435.5629	0.8569	0.8282
0.5979	0.58	4400	0.5594	-1.6981	-2.4801	0.7040	0.7820	-511.8796	-455.1061	0.9415	0.9060
0.4859	0.59	4500	0.5530	-1.5910	-2.3517	0.7080	0.7607	-499.0415	-444.3948	0.9399	0.9057
0.5484	0.6	4600	0.5525	-1.5159	-2.2439	0.7055	0.7280	-488.2595	-436.8822	0.8711	0.8268
0.6135	0.62	4700	0.5504	-1.3255	-2.0246	0.7065	0.6990	-466.3248	-417.8462	0.7736	0.7222
0.5714	0.63	4800	0.5501	-1.4736	-2.1670	0.7070	0.6934	-480.5717	-432.6558	0.8649	0.8370
0.517	0.64	4900	0.5531	-1.6509	-2.4069	0.7090	0.7560	-504.5561	-450.3797	0.9735	0.9524
0.4862	0.65	5000	0.5524	-1.5409	-2.2932	0.7080	0.7523	-493.1930	-439.3873	0.9138	0.8849
0.6176	0.67	5100	0.5519	-1.4759	-2.2276	0.7020	0.7516	-486.6266	-432.8859	0.8785	0.8443
0.5514	0.68	5200	0.5500	-1.4083	-2.1357	0.7025	0.7274	-477.4418	-426.1200	0.8299	0.7894
0.5166	0.69	5300	0.5508	-1.4154	-2.1510	0.7040	0.7356	-478.9723	-426.8324	0.8441	0.8065
0.4918	0.71	5400	0.5496	-1.4093	-2.1290	0.7090	0.7197	-476.7667	-426.2183	0.8313	0.7905
0.596	0.72	5500	0.5489	-1.4890	-2.2221	0.7075	0.7332	-486.0821	-434.1885	0.8632	0.8239
0.6034	0.73	5600	0.5489	-1.4048	-2.1338	0.7065	0.7290	-477.2522	-425.7730	0.8041	0.7561
0.4793	0.75	5700	0.5495	-1.5017	-2.2541	0.7080	0.7524	-489.2809	-435.4676	0.8918	0.8545
0.5164	0.76	5800	0.5497	-1.5548	-2.3215	0.7085	0.7667	-496.0150	-440.7685	0.9221	0.8885
0.6164	0.77	5900	0.5491	-1.5335	-2.2884	0.7080	0.7549	-492.7101	-438.6432	0.8987	0.8645
0.5347	0.79	6000	0.5487	-1.5028	-2.2487	0.7105	0.7459	-488.7427	-435.5721	0.8766	0.8397
0.56	0.8	6100	0.5491	-1.4855	-2.2337	0.7105	0.7482	-487.2426	-433.8429	0.8643	0.8248
0.587	0.81	6200	0.5491	-1.4638	-2.2111	0.7095	0.7473	-484.9788	-431.6711	0.8489	0.8072
0.4927	0.82	6300	0.5490	-1.4591	-2.2082	0.7090	0.7491	-484.6881	-431.2039	0.8531	0.8118
0.6102	0.84	6400	0.5486	-1.4462	-2.1928	0.7105	0.7466	-483.1518	-429.9117	0.8474	0.8055
0.4988	0.85	6500	0.5485	-1.4482	-2.1938	0.7095	0.7456	-483.2466	-430.1142	0.8464	0.8046
0.5544	0.86	6600	0.5486	-1.4491	-2.1949	0.7115	0.7458	-483.3600	-430.1988	0.8487	0.8068
0.5828	0.88	6700	0.5486	-1.4518	-2.1981	0.7100	0.7463	-483.6802	-430.4771	0.8512	0.8097
0.5711	0.89	6800	0.5485	-1.4557	-2.2030	0.7095	0.7473	-484.1660	-430.8610	0.8538	0.8124
0.5621	0.9	6900	0.5484	-1.4557	-2.2035	0.7125	0.7478	-484.2229	-430.8625	0.8535	0.8119
0.5093	0.92	7000	0.5485	-1.4555	-2.2030	0.7095	0.7475	-484.1658	-430.8411	0.8539	0.8128
0.4665	0.93	7100	0.5485	-1.4561	-2.2038	0.7100	0.7477	-484.2509	-430.9035	0.8539	0.8128
0.6276	0.94	7200	0.5486	-1.4556	-2.2033	0.7110	0.7476	-484.1955	-430.8554	0.8539	0.8130
0.457	0.96	7300	0.5486	-1.4547	-2.2022	0.7110	0.7475	-484.0942	-430.7640	0.8540	0.8129
0.5436	0.97	7400	0.5486	-1.4557	-2.2035	0.7130	0.7478	-484.2209	-430.8634	0.8541	0.8130
0.4801	0.98	7500	0.5486	-1.4555	-2.2033	0.7125	0.7478	-484.1994	-430.8404	0.8538	0.8125
0.5922	0.99	7600	0.5486	-1.4555	-2.2032	0.7100	0.7477	-484.1860	-430.8414	0.8537	0.8124

Framework versions

• PEFT 0.7.1
• Transformers 4.36.2
• Pytorch 2.1.2
• Datasets 2.14.6
• Tokenizers 0.15.0