Ministral-8B-Instruct-2410-dpo-1000

This model is a fine-tuned version of mistralai/Ministral-8B-Instruct-2410 on the bct_non_cot_dpo_1000 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3134
• Rewards/chosen: -0.2081
• Rewards/rejected: -2.0644
• Rewards/accuracies: 0.9000
• Rewards/margins: 1.8564
• Logps/chosen: -27.0565
• Logps/rejected: -47.3938
• Logits/chosen: -1.2929
• Logits/rejected: -1.3581

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
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/chosen	Logps/rejected	Logits/chosen	Logits/rejected
0.5697	1.7778	50	0.5310	0.0503	-0.3398	0.8400	0.3901	-24.4728	-30.1476	-1.6708	-1.7014
0.3795	3.5556	100	0.3455	-0.1007	-1.3901	0.9000	1.2894	-25.9826	-40.6504	-1.4192	-1.4765
0.2291	5.3333	150	0.3134	-0.2081	-2.0644	0.9000	1.8564	-27.0565	-47.3938	-1.2929	-1.3581
0.2516	7.1111	200	0.3151	-0.2309	-2.3286	0.8900	2.0977	-27.2845	-50.0355	-1.2557	-1.3219
0.1897	8.8889	250	0.3143	-0.2393	-2.3851	0.8900	2.1459	-27.3683	-50.6008	-1.2463	-1.3122

Framework versions

• PEFT 0.12.0
• Transformers 4.45.2
• Pytorch 2.3.0
• Datasets 2.19.0
• Tokenizers 0.20.0