martimfasantos's picture
End of training
624245f verified
|
raw
history blame
13.2 kB
metadata
license: apache-2.0
library_name: peft
tags:
  - alignment-handbook
  - trl
  - dpo
  - generated_from_trainer
base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
datasets:
  - openai/summarize_from_feedback
model-index:
  - name: tinyllama-1.1b-sum-dpo-qlora
    results: []

tinyllama-1.1b-sum-dpo-qlora

This model is a fine-tuned version of martimfasantos/tinyllama-1.1b-sum-sft-qlora on the openai/summarize_from_feedback dataset. It achieves the following results on the evaluation set:

  • Loss: 0.6463
  • Rewards/chosen: -0.9560
  • Rewards/rejected: -1.1279
  • Rewards/accuracies: 0.6204
  • Rewards/margins: 0.1719
  • Logps/rejected: -187.9012
  • Logps/chosen: -167.0102
  • Logits/rejected: -3.0162
  • Logits/chosen: -3.0224

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: 8
  • seed: 42
  • distributed_type: multi-GPU
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 16
  • 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
0.6927 0.02 100 0.6930 0.0049 0.0047 0.5204 0.0003 -74.6416 -70.9175 -3.4966 -3.4983
0.692 0.03 200 0.6926 0.0146 0.0135 0.5616 0.0012 -73.7585 -69.9458 -3.4924 -3.4942
0.6887 0.05 300 0.6911 0.0351 0.0308 0.5732 0.0043 -72.0302 -67.9024 -3.4858 -3.4876
0.6865 0.07 400 0.6890 0.0164 0.0077 0.5609 0.0087 -74.3370 -69.7677 -3.4786 -3.4805
0.6864 0.09 500 0.6864 0.0236 0.0089 0.5755 0.0146 -74.2129 -69.0538 -3.4662 -3.4680
0.6731 0.1 600 0.6838 0.0019 -0.0189 0.5871 0.0209 -77.0012 -71.2189 -3.4497 -3.4515
0.6749 0.12 700 0.6788 -0.0758 -0.1091 0.5980 0.0333 -86.0178 -78.9945 -3.4470 -3.4489
0.6678 0.14 800 0.6741 -0.1859 -0.2330 0.5906 0.0471 -98.4033 -89.9991 -3.4169 -3.4188
0.6655 0.16 900 0.6709 -0.1856 -0.2411 0.5927 0.0555 -99.2188 -89.9669 -3.3811 -3.3826
0.6695 0.17 1000 0.6686 -0.3893 -0.4584 0.5946 0.0691 -120.9453 -110.3432 -3.3595 -3.3611
0.6648 0.19 1100 0.6702 -0.2078 -0.2671 0.5976 0.0593 -101.8174 -92.1903 -3.3439 -3.3453
0.6543 0.21 1200 0.6642 -0.3511 -0.4313 0.6011 0.0802 -118.2354 -106.5216 -3.3096 -3.3110
0.6535 0.22 1300 0.6605 -0.4651 -0.5609 0.5990 0.0957 -131.1967 -117.9248 -3.2817 -3.2832
0.6315 0.24 1400 0.6606 -0.3801 -0.4704 0.6138 0.0903 -122.1497 -109.4246 -3.2773 -3.2788
0.6595 0.26 1500 0.6544 -0.5561 -0.6712 0.6197 0.1151 -142.2231 -127.0196 -3.2429 -3.2446
0.6383 0.28 1600 0.6538 -0.5868 -0.7052 0.6178 0.1184 -145.6309 -130.0926 -3.2318 -3.2338
0.6775 0.29 1700 0.6568 -0.4687 -0.5717 0.6173 0.1030 -132.2748 -118.2820 -3.2194 -3.2212
0.6312 0.31 1800 0.6497 -0.7203 -0.8617 0.6111 0.1414 -161.2767 -143.4406 -3.1213 -3.1237
0.665 0.33 1900 0.6551 -0.5175 -0.6278 0.6134 0.1103 -137.8867 -123.1614 -3.1660 -3.1680
0.6385 0.34 2000 0.6522 -0.6166 -0.7379 0.6162 0.1213 -148.8959 -133.0700 -3.1823 -3.1845
0.6452 0.36 2100 0.6538 -0.7088 -0.8325 0.6048 0.1237 -158.3535 -142.2912 -3.1344 -3.1369
0.6024 0.38 2200 0.6527 -0.6378 -0.7639 0.6120 0.1262 -151.5019 -135.1858 -3.1567 -3.1596
0.5912 0.4 2300 0.6485 -0.8992 -1.0561 0.6106 0.1569 -180.7164 -161.3302 -3.0812 -3.0853
0.6188 0.41 2400 0.6488 -0.9960 -1.1662 0.6204 0.1702 -191.7268 -171.0100 -3.0219 -3.0276
0.6286 0.43 2500 0.6483 -0.8764 -1.0333 0.6076 0.1568 -178.4354 -159.0542 -3.0428 -3.0475
0.61 0.45 2600 0.6532 -0.7428 -0.8730 0.6018 0.1302 -162.4074 -145.6894 -3.0767 -3.0804
0.6295 0.47 2700 0.6526 -0.6786 -0.8083 0.6138 0.1296 -155.9322 -139.2748 -3.1080 -3.1114
0.6504 0.48 2800 0.6510 -0.7810 -0.9243 0.6106 0.1432 -167.5323 -149.5115 -3.0877 -3.0915
0.6226 0.5 2900 0.6513 -0.7637 -0.9050 0.6127 0.1413 -165.6116 -147.7837 -3.0831 -3.0870
0.6226 0.52 3000 0.6494 -0.7375 -0.8834 0.6078 0.1459 -163.4444 -145.1619 -3.0916 -3.0955
0.6062 0.53 3100 0.6485 -0.7793 -0.9311 0.6129 0.1518 -168.2215 -149.3398 -3.0906 -3.0949
0.6071 0.55 3200 0.6477 -0.8041 -0.9577 0.6118 0.1536 -170.8775 -151.8242 -3.0911 -3.0956
0.608 0.57 3300 0.6461 -1.1115 -1.2974 0.6150 0.1859 -204.8467 -182.5597 -3.0002 -3.0064
0.5996 0.59 3400 0.6486 -0.7960 -0.9481 0.6099 0.1520 -169.9129 -151.0113 -3.0691 -3.0742
0.6081 0.6 3500 0.6478 -0.8354 -0.9930 0.6157 0.1576 -174.4116 -154.9542 -3.0630 -3.0681
0.6256 0.62 3600 0.6491 -0.7744 -0.9234 0.6145 0.1489 -167.4422 -148.8546 -3.0722 -3.0769
0.5969 0.64 3700 0.6469 -0.9732 -1.1419 0.6150 0.1687 -189.2978 -168.7282 -3.0171 -3.0231
0.6272 0.65 3800 0.6472 -0.9477 -1.1124 0.6176 0.1648 -186.3489 -166.1768 -3.0087 -3.0145
0.6222 0.67 3900 0.6467 -0.9719 -1.1400 0.6166 0.1681 -189.1107 -168.6043 -3.0040 -3.0100
0.605 0.69 4000 0.6461 -1.0773 -1.2558 0.6204 0.1785 -200.6857 -179.1379 -2.9783 -2.9849
0.585 0.71 4100 0.6464 -0.9836 -1.1556 0.6164 0.1720 -190.6670 -169.7659 -3.0024 -3.0086
0.6602 0.72 4200 0.6465 -0.9496 -1.1182 0.6178 0.1686 -186.9268 -166.3669 -3.0089 -3.0150
0.6074 0.74 4300 0.6468 -0.8954 -1.0597 0.6183 0.1643 -181.0816 -160.9504 -3.0248 -3.0306
0.6105 0.76 4400 0.6470 -0.8905 -1.0547 0.6150 0.1641 -180.5745 -160.4626 -3.0306 -3.0365
0.6127 0.78 4500 0.6470 -0.8899 -1.0538 0.6183 0.1638 -180.4842 -160.4037 -3.0280 -3.0338
0.5798 0.79 4600 0.6468 -0.9128 -1.0793 0.6208 0.1665 -183.0344 -162.6864 -3.0195 -3.0255
0.6228 0.81 4700 0.6467 -0.9215 -1.0896 0.6192 0.1681 -184.0640 -163.5562 -3.0231 -3.0291
0.6131 0.83 4800 0.6466 -0.9391 -1.1091 0.6199 0.1700 -186.0176 -165.3165 -3.0141 -3.0202
0.6215 0.84 4900 0.6465 -0.9478 -1.1189 0.6197 0.1711 -186.9947 -166.1919 -3.0180 -3.0241
0.585 0.86 5000 0.6460 -0.9592 -1.1321 0.6201 0.1729 -188.3154 -167.3252 -3.0164 -3.0226
0.6478 0.88 5100 0.6460 -0.9606 -1.1336 0.6194 0.1730 -188.4695 -167.4737 -3.0151 -3.0213
0.6018 0.9 5200 0.6462 -0.9572 -1.1296 0.6206 0.1725 -188.0692 -167.1259 -3.0105 -3.0167
0.5963 0.91 5300 0.6465 -0.9564 -1.1282 0.6199 0.1718 -187.9285 -167.0541 -3.0167 -3.0229
0.5921 0.93 5400 0.6462 -0.9569 -1.1292 0.6199 0.1723 -188.0274 -167.0996 -3.0133 -3.0196
0.6015 0.95 5500 0.6463 -0.9570 -1.1292 0.6192 0.1723 -188.0282 -167.1056 -3.0164 -3.0226
0.6148 0.96 5600 0.6461 -0.9543 -1.1269 0.6194 0.1726 -187.7934 -166.8396 -3.0142 -3.0205
0.6299 0.98 5700 0.6462 -0.9543 -1.1263 0.6194 0.1720 -187.7363 -166.8363 -3.0166 -3.0228
0.5854 1.0 5800 0.6463 -0.9560 -1.1279 0.6204 0.1719 -187.9012 -167.0102 -3.0162 -3.0224

Framework versions

  • PEFT 0.7.1
  • Transformers 4.39.3
  • Pytorch 2.1.2
  • Datasets 2.18.0
  • Tokenizers 0.15.2