|
--- |
|
base_model: alignment-handbook/zephyr-7b-sft-full |
|
datasets: |
|
- generation/UF |
|
library_name: peft |
|
license: apache-2.0 |
|
tags: |
|
- alignment-handbook |
|
- trl |
|
- dpo |
|
- generated_from_trainer |
|
model-index: |
|
- name: zephyr-dpo-qlora-uf-ours-5e-7 |
|
results: [] |
|
--- |
|
|
|
<!-- This model card has been generated automatically according to the information the Trainer had access to. You |
|
should probably proofread and complete it, then remove this comment. --> |
|
|
|
# zephyr-dpo-qlora-uf-ours-5e-7 |
|
|
|
This model is a fine-tuned version of [alignment-handbook/zephyr-7b-sft-full](https://huggingface.co/alignment-handbook/zephyr-7b-sft-full) on the generation/UF dataset. |
|
It achieves the following results on the evaluation set: |
|
- Loss: 0.6585 |
|
- Rewards/chosen: -0.1318 |
|
- Rewards/rejected: -0.2668 |
|
- Rewards/accuracies: 0.6360 |
|
- Rewards/margins: 0.1350 |
|
- Rewards/margins Max: 0.7123 |
|
- Rewards/margins Min: -0.4480 |
|
- Rewards/margins Std: 0.4011 |
|
- Logps/rejected: -285.2624 |
|
- Logps/chosen: -297.7779 |
|
- Logits/rejected: -2.6359 |
|
- Logits/chosen: -2.6686 |
|
|
|
## 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: 4 |
|
- eval_batch_size: 8 |
|
- seed: 42 |
|
- distributed_type: multi-GPU |
|
- num_devices: 2 |
|
- gradient_accumulation_steps: 2 |
|
- total_train_batch_size: 16 |
|
- total_eval_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: 3 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Rewards/margins Max | Rewards/margins Min | Rewards/margins Std | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen | |
|
|:-------------:|:-----:|:----:|:---------------:|:--------------:|:----------------:|:------------------:|:---------------:|:-------------------:|:-------------------:|:-------------------:|:--------------:|:------------:|:---------------:|:-------------:| |
|
| 0.6872 | 0.28 | 100 | 0.6920 | 0.0070 | 0.0040 | 0.5890 | 0.0030 | 0.0201 | -0.0117 | 0.0106 | -258.1791 | -283.8921 | -2.7682 | -2.8069 | |
|
| 0.6585 | 0.56 | 200 | 0.6870 | 0.0392 | 0.0222 | 0.5990 | 0.0171 | 0.1011 | -0.0563 | 0.0523 | -256.3612 | -280.6689 | -2.7497 | -2.7877 | |
|
| 0.6099 | 0.85 | 300 | 0.6805 | 0.0140 | -0.0224 | 0.5980 | 0.0364 | 0.2144 | -0.1227 | 0.1127 | -260.8226 | -283.1934 | -2.7177 | -2.7552 | |
|
| 0.5685 | 1.13 | 400 | 0.6739 | -0.0306 | -0.0901 | 0.6020 | 0.0596 | 0.3565 | -0.2113 | 0.1904 | -267.5928 | -287.6518 | -2.6871 | -2.7235 | |
|
| 0.5275 | 1.41 | 500 | 0.6691 | -0.0498 | -0.1325 | 0.6110 | 0.0827 | 0.4946 | -0.3139 | 0.2729 | -271.8259 | -289.5688 | -2.6519 | -2.6881 | |
|
| 0.4992 | 1.69 | 600 | 0.6633 | -0.0820 | -0.1909 | 0.6240 | 0.1089 | 0.6097 | -0.3758 | 0.3364 | -277.6682 | -292.7885 | -2.6447 | -2.6793 | |
|
| 0.4721 | 1.97 | 700 | 0.6607 | -0.0596 | -0.1836 | 0.6320 | 0.1240 | 0.6617 | -0.3928 | 0.3590 | -276.9363 | -290.5549 | -2.6452 | -2.6783 | |
|
| 0.4401 | 2.25 | 800 | 0.6595 | -0.1036 | -0.2352 | 0.6340 | 0.1316 | 0.6923 | -0.4292 | 0.3871 | -282.0996 | -294.9506 | -2.6407 | -2.6735 | |
|
| 0.4402 | 2.54 | 900 | 0.6586 | -0.1091 | -0.2452 | 0.6450 | 0.1361 | 0.7080 | -0.4378 | 0.3951 | -283.1014 | -295.5010 | -2.6372 | -2.6699 | |
|
| 0.4835 | 2.82 | 1000 | 0.6584 | -0.1313 | -0.2662 | 0.6380 | 0.1349 | 0.7122 | -0.4482 | 0.4013 | -285.1998 | -297.7245 | -2.6391 | -2.6714 | |
|
|
|
|
|
### Framework versions |
|
|
|
- PEFT 0.7.1 |
|
- Transformers 4.39.0.dev0 |
|
- Pytorch 2.1.2+cu121 |
|
- Datasets 2.14.6 |
|
- Tokenizers 0.15.2 |
