Edit model card

phi_1_5_dpo_ep6

This model is a fine-tuned version of /home/work/saic-llm-2023/checkpoints/microsoft/phi-1_5 on the argilla/ultrafeedback-binarized-preferences-cleaned dataset. It achieves the following results on the evaluation set:

  • Loss: 0.4748
  • Rewards/chosen: -0.9135
  • Rewards/rejected: -1.9448
  • Rewards/accuracies: 0.7937
  • Rewards/margins: 1.0313
  • Logps/rejected: -618.5530
  • Logps/chosen: -634.6866
  • Logits/rejected: 3.4318
  • Logits/chosen: 3.4052

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-07
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 4
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 64
  • total_eval_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_steps: 100
  • num_epochs: 6

Training results

Training Loss Epoch Step Validation Loss Rewards/chosen Rewards/rejected Rewards/accuracies Rewards/margins Logps/rejected Logps/chosen Logits/rejected Logits/chosen
0.6881 0.11 100 0.6856 0.0468 0.0298 0.7024 0.0170 -421.0949 -538.6564 4.8883 4.6646
0.6692 0.22 200 0.6642 0.1742 0.0988 0.7123 0.0754 -414.1955 -525.9189 4.8718 4.6370
0.6368 0.33 300 0.6442 0.2557 0.1261 0.7083 0.1296 -411.4657 -517.7680 4.8407 4.5968
0.6283 0.43 400 0.6283 0.2608 0.0812 0.7083 0.1795 -415.9522 -517.2609 4.7629 4.5156
0.6052 0.54 500 0.6132 0.1429 -0.0998 0.7103 0.2427 -434.0545 -529.0491 4.5516 4.3153
0.5923 0.65 600 0.6008 0.1425 -0.1628 0.7123 0.3053 -440.3539 -529.0887 4.4588 4.2289
0.5899 0.76 700 0.5880 0.0755 -0.2915 0.7083 0.3670 -453.2271 -535.7857 4.3444 4.1349
0.558 0.87 800 0.5715 -0.0965 -0.5304 0.7262 0.4339 -477.1144 -552.9822 4.2704 4.0642
0.5495 0.98 900 0.5552 -0.2658 -0.7677 0.7341 0.5019 -500.8484 -569.9210 4.1976 4.0015
0.5124 1.09 1000 0.5473 -0.3871 -0.9394 0.7321 0.5523 -518.0129 -582.0427 4.0959 3.9125
0.5322 1.19 1100 0.5400 -0.3641 -0.9463 0.7579 0.5821 -518.7011 -579.7518 4.0436 3.8715
0.5281 1.3 1200 0.5344 -0.5340 -1.1498 0.7460 0.6158 -539.0579 -596.7365 3.9368 3.7842
0.5063 1.41 1300 0.5297 -0.3754 -0.9975 0.7579 0.6221 -523.8221 -580.8731 4.0135 3.8499
0.5073 1.52 1400 0.5216 -0.3819 -1.0300 0.7758 0.6481 -527.0738 -581.5236 3.9401 3.7846
0.5156 1.63 1500 0.5177 -0.5748 -1.2824 0.7560 0.7077 -552.3166 -600.8123 3.7868 3.6678
0.5072 1.74 1600 0.5138 -0.4973 -1.2122 0.7798 0.7149 -545.2914 -593.0637 3.7791 3.6614
0.4908 1.85 1700 0.5077 -0.5479 -1.2972 0.7798 0.7493 -553.7918 -598.1292 3.7893 3.6696
0.5109 1.95 1800 0.5068 -0.6157 -1.3930 0.7758 0.7773 -563.3733 -604.9089 3.7679 3.6556
0.4779 2.06 1900 0.5005 -0.6247 -1.4169 0.7738 0.7922 -565.7673 -605.8088 3.7118 3.6062
0.4833 2.17 2000 0.4992 -0.6841 -1.5026 0.7698 0.8185 -574.3334 -611.7432 3.6739 3.5849
0.4879 2.28 2100 0.4967 -0.8128 -1.6654 0.7698 0.8526 -590.6146 -624.6127 3.5692 3.5030
0.4645 2.39 2200 0.4927 -0.6969 -1.5365 0.7857 0.8396 -577.7230 -613.0289 3.6647 3.5772
0.4587 2.5 2300 0.4936 -0.6024 -1.4533 0.7778 0.8509 -569.4068 -603.5743 3.6615 3.5790
0.437 2.61 2400 0.4921 -0.8826 -1.7724 0.7738 0.8897 -601.3099 -631.5984 3.4903 3.4343
0.4204 2.71 2500 0.4890 -0.8338 -1.7338 0.7758 0.8999 -597.4498 -626.7175 3.5447 3.4804
0.467 2.82 2600 0.4865 -0.5910 -1.4516 0.7877 0.8606 -569.2333 -602.4326 3.5690 3.5000
0.458 2.93 2700 0.4861 -0.7666 -1.6726 0.7837 0.9059 -591.3298 -620.0014 3.5208 3.4579
0.462 3.04 2800 0.4844 -0.7109 -1.6145 0.7917 0.9037 -585.5269 -614.4227 3.5553 3.4954
0.4258 3.15 2900 0.4888 -0.9814 -1.9414 0.7817 0.9600 -618.2142 -641.4772 3.4761 3.4227
0.4219 3.26 3000 0.4856 -0.8858 -1.8323 0.7937 0.9465 -607.3071 -631.9181 3.4895 3.4362
0.4295 3.37 3100 0.4823 -0.8140 -1.7651 0.7976 0.9511 -600.5797 -624.7327 3.4880 3.4357
0.4268 3.47 3200 0.4800 -0.8592 -1.8282 0.7976 0.9690 -606.8929 -629.2567 3.4536 3.4126
0.4338 3.58 3300 0.4785 -0.8784 -1.8458 0.7956 0.9674 -608.6551 -631.1731 3.4471 3.4096
0.4297 3.69 3400 0.4774 -0.9026 -1.8929 0.7956 0.9903 -613.3634 -633.5962 3.4710 3.4326
0.4133 3.8 3500 0.4785 -0.9173 -1.9072 0.7937 0.9899 -614.7964 -635.0674 3.4610 3.4232
0.4275 3.91 3600 0.4794 -1.0209 -2.0380 0.7837 1.0171 -627.8748 -645.4227 3.4635 3.4227
0.4224 4.02 3700 0.4784 -0.9130 -1.9086 0.7937 0.9955 -614.9320 -634.6396 3.4812 3.4400
0.4101 4.13 3800 0.4773 -0.9474 -1.9571 0.7877 1.0097 -619.7819 -638.0772 3.4569 3.4225
0.4295 4.23 3900 0.4790 -0.9893 -2.0096 0.7956 1.0203 -625.0361 -642.2666 3.4290 3.3998
0.4162 4.34 4000 0.4769 -0.9682 -1.9897 0.7956 1.0215 -623.0465 -640.1562 3.4342 3.4040
0.425 4.45 4100 0.4759 -0.9553 -1.9788 0.7917 1.0236 -621.9555 -638.8621 3.4580 3.4237
0.4155 4.56 4200 0.4778 -1.0183 -2.0573 0.7917 1.0390 -629.8077 -645.1696 3.4277 3.3981
0.4311 4.67 4300 0.4765 -0.9712 -2.0065 0.7897 1.0353 -624.7266 -640.4598 3.4413 3.4107
0.41 4.78 4400 0.4768 -0.9764 -2.0101 0.7917 1.0337 -625.0818 -640.9733 3.4387 3.4081
0.4127 4.89 4500 0.4749 -0.9599 -1.9994 0.7937 1.0395 -624.0168 -639.3277 3.4453 3.4160
0.453 4.99 4600 0.4748 -0.9231 -1.9528 0.7917 1.0297 -619.3519 -635.6462 3.4444 3.4142
0.4035 5.1 4700 0.4754 -0.9561 -1.9965 0.7897 1.0403 -623.7211 -638.9504 3.4293 3.4019
0.4225 5.21 4800 0.4753 -0.9471 -1.9855 0.7877 1.0384 -622.6226 -638.0461 3.4359 3.4077
0.3941 5.32 4900 0.4754 -0.9579 -1.9978 0.7897 1.0400 -623.8593 -639.1230 3.4282 3.4012
0.4093 5.43 5000 0.4748 -0.9135 -1.9448 0.7937 1.0313 -618.5530 -634.6866 3.4318 3.4052
0.3902 5.54 5100 0.4754 -0.9457 -1.9815 0.7956 1.0358 -622.2274 -637.9056 3.4281 3.4014
0.3795 5.65 5200 0.4753 -0.9484 -1.9852 0.7897 1.0368 -622.5895 -638.1724 3.4253 3.3988
0.3915 5.75 5300 0.4754 -0.9571 -1.9957 0.7956 1.0386 -623.6450 -639.0427 3.4242 3.3979
0.4075 5.86 5400 0.4756 -0.9566 -1.9949 0.7877 1.0383 -623.5674 -638.9974 3.4221 3.3962
0.4293 5.97 5500 0.4756 -0.9571 -1.9948 0.7897 1.0377 -623.5548 -639.0446 3.4230 3.3964

Framework versions

  • Transformers 4.38.0
  • Pytorch 2.1.2+cu118
  • Datasets 2.17.1
  • Tokenizers 0.15.0
Downloads last month
4
Safetensors
Model size
1.42B params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for ondevicellm/phi_1_5_dpo_ep6

Base model

microsoft/phi-1_5
Finetuned
(220)
this model

Dataset used to train ondevicellm/phi_1_5_dpo_ep6