MedQA_L3_1000steps_1e7rate_01beta_CSFTDPO

This model is a fine-tuned version of tsavage68/MedQA_L3_1000steps_1e6rate_SFT on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6556
• Rewards/chosen: 0.3104
• Rewards/rejected: 0.2288
• Rewards/accuracies: 0.7187
• Rewards/margins: 0.0816
• Logps/rejected: -31.5670
• Logps/chosen: -28.2248
• Logits/rejected: -0.7354
• Logits/chosen: -0.7346

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-07
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• 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
0.6932	0.0489	50	0.6927	-0.0017	-0.0025	0.5297	0.0008	-33.8801	-31.3453	-0.7320	-0.7313
0.691	0.0977	100	0.6894	0.0852	0.0776	0.6505	0.0076	-33.0791	-30.4769	-0.7328	-0.7321
0.6779	0.1466	150	0.6824	0.1496	0.1271	0.6791	0.0225	-32.5836	-29.8325	-0.7314	-0.7307
0.6695	0.1954	200	0.6773	0.0689	0.0354	0.6945	0.0335	-33.5008	-30.6395	-0.7313	-0.7306
0.6792	0.2443	250	0.6730	0.1279	0.0855	0.7231	0.0424	-32.9998	-30.0495	-0.7313	-0.7306
0.6641	0.2931	300	0.6678	0.1588	0.1052	0.7297	0.0536	-32.8025	-29.7403	-0.7323	-0.7315
0.665	0.3420	350	0.6652	0.2014	0.1419	0.7187	0.0595	-32.4354	-29.3144	-0.7344	-0.7336
0.6504	0.3908	400	0.6621	0.2655	0.1993	0.7363	0.0662	-31.8619	-28.6732	-0.7340	-0.7332
0.6533	0.4397	450	0.6607	0.2838	0.2142	0.7319	0.0697	-31.7132	-28.4903	-0.7347	-0.7339
0.66	0.4885	500	0.6588	0.3223	0.2481	0.7187	0.0742	-31.3734	-28.1056	-0.7350	-0.7342
0.6373	0.5374	550	0.6578	0.3176	0.2410	0.7143	0.0766	-31.4445	-28.1521	-0.7355	-0.7347
0.6608	0.5862	600	0.6566	0.3164	0.2373	0.7187	0.0792	-31.4823	-28.1640	-0.7357	-0.7349
0.6457	0.6351	650	0.6560	0.3040	0.2233	0.7187	0.0807	-31.6215	-28.2882	-0.7350	-0.7342
0.657	0.6839	700	0.6554	0.3088	0.2267	0.7165	0.0820	-31.5874	-28.2407	-0.7349	-0.7341
0.6597	0.7328	750	0.6560	0.3104	0.2296	0.7187	0.0808	-31.5590	-28.2246	-0.7355	-0.7346
0.6642	0.7816	800	0.6553	0.3115	0.2291	0.7209	0.0824	-31.5639	-28.2138	-0.7353	-0.7345
0.673	0.8305	850	0.6555	0.3114	0.2296	0.7231	0.0818	-31.5592	-28.2146	-0.7352	-0.7344
0.6659	0.8793	900	0.6556	0.3142	0.2324	0.7143	0.0818	-31.5308	-28.1868	-0.7357	-0.7349
0.6533	0.9282	950	0.6556	0.3104	0.2288	0.7187	0.0816	-31.5668	-28.2246	-0.7354	-0.7346
0.6255	0.9770	1000	0.6556	0.3104	0.2288	0.7187	0.0816	-31.5670	-28.2248	-0.7354	-0.7346

Framework versions

• Transformers 4.41.1
• Pytorch 2.0.0+cu117
• Datasets 2.19.1
• Tokenizers 0.19.1