--- license: apache-2.0 base_model: mosaicml/mpt-7b-instruct tags: - trl - dpo - generated_from_trainer model-index: - name: MPT_1000_STEPS_1e8_rate_01_beta_DPO results: [] --- # MPT_1000_STEPS_1e8_rate_01_beta_DPO This model is a fine-tuned version of [mosaicml/mpt-7b-instruct](https://huggingface.co/mosaicml/mpt-7b-instruct) on an unknown dataset. It achieves the following results on the evaluation set: - Loss: 0.6939 - Rewards/chosen: -0.0008 - Rewards/rejected: 0.0005 - Rewards/accuracies: 0.4747 - Rewards/margins: -0.0013 - Logps/rejected: -21.5525 - Logps/chosen: -20.8004 - Logits/rejected: 14.2517 - Logits/chosen: 14.2543 ## 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-08 - 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 | |:-------------:|:-----:|:----:|:---------------:|:--------------:|:----------------:|:------------------:|:---------------:|:--------------:|:------------:|:---------------:|:-------------:| | 0.6939 | 0.05 | 50 | 0.6936 | -0.0008 | -0.0000 | 0.4791 | -0.0007 | -21.5578 | -20.7998 | 14.2549 | 14.2575 | | 0.6946 | 0.1 | 100 | 0.6930 | -0.0001 | -0.0005 | 0.4923 | 0.0004 | -21.5626 | -20.7929 | 14.2554 | 14.2579 | | 0.6931 | 0.15 | 150 | 0.6935 | -0.0014 | -0.0009 | 0.4967 | -0.0005 | -21.5666 | -20.8066 | 14.2572 | 14.2598 | | 0.6917 | 0.2 | 200 | 0.6929 | 0.0003 | -0.0004 | 0.4813 | 0.0006 | -21.5611 | -20.7895 | 14.2562 | 14.2588 | | 0.6954 | 0.24 | 250 | 0.6940 | -0.0021 | -0.0006 | 0.4857 | -0.0015 | -21.5632 | -20.8129 | 14.2623 | 14.2649 | | 0.6932 | 0.29 | 300 | 0.6931 | -0.0008 | -0.0009 | 0.4967 | 0.0001 | -21.5667 | -20.8001 | 14.2610 | 14.2636 | | 0.6954 | 0.34 | 350 | 0.6934 | -0.0012 | -0.0009 | 0.5011 | -0.0003 | -21.5662 | -20.8041 | 14.2641 | 14.2667 | | 0.6891 | 0.39 | 400 | 0.6945 | -0.0025 | 0.0001 | 0.4725 | -0.0026 | -21.5566 | -20.8174 | 14.2546 | 14.2572 | | 0.6917 | 0.44 | 450 | 0.6935 | -0.0011 | -0.0005 | 0.4593 | -0.0006 | -21.5622 | -20.8030 | 14.2562 | 14.2588 | | 0.6908 | 0.49 | 500 | 0.6936 | -0.0018 | -0.0010 | 0.4813 | -0.0007 | -21.5679 | -20.8101 | 14.2507 | 14.2533 | | 0.6927 | 0.54 | 550 | 0.6941 | -0.0016 | 0.0001 | 0.4681 | -0.0017 | -21.5560 | -20.8079 | 14.2549 | 14.2575 | | 0.6923 | 0.59 | 600 | 0.6939 | -0.0011 | 0.0003 | 0.4527 | -0.0014 | -21.5542 | -20.8035 | 14.2539 | 14.2565 | | 0.6946 | 0.64 | 650 | 0.6944 | -0.0013 | 0.0012 | 0.4593 | -0.0024 | -21.5459 | -20.8048 | 14.2527 | 14.2553 | | 0.6918 | 0.68 | 700 | 0.6934 | -0.0002 | 0.0002 | 0.4747 | -0.0004 | -21.5558 | -20.7942 | 14.2531 | 14.2557 | | 0.6923 | 0.73 | 750 | 0.6939 | -0.0011 | 0.0002 | 0.4967 | -0.0013 | -21.5551 | -20.8028 | 14.2522 | 14.2547 | | 0.6895 | 0.78 | 800 | 0.6937 | -0.0006 | 0.0004 | 0.4945 | -0.0010 | -21.5532 | -20.7977 | 14.2513 | 14.2539 | | 0.6936 | 0.83 | 850 | 0.6937 | -0.0007 | 0.0003 | 0.4945 | -0.0010 | -21.5541 | -20.7990 | 14.2516 | 14.2542 | | 0.6904 | 0.88 | 900 | 0.6939 | -0.0008 | 0.0005 | 0.4747 | -0.0013 | -21.5525 | -20.8004 | 14.2517 | 14.2543 | | 0.6926 | 0.93 | 950 | 0.6939 | -0.0008 | 0.0005 | 0.4747 | -0.0013 | -21.5525 | -20.8004 | 14.2517 | 14.2543 | | 0.691 | 0.98 | 1000 | 0.6939 | -0.0008 | 0.0005 | 0.4747 | -0.0013 | -21.5525 | -20.8004 | 14.2517 | 14.2543 | ### Framework versions - Transformers 4.39.1 - Pytorch 2.0.0+cu117 - Datasets 2.18.0 - Tokenizers 0.15.2