metadata
license: apache-2.0
tags:
- generated_from_trainer
base_model: amazingvince/zephyr-smol_llama-100m-sft-full
model-index:
- name: zephyr-smol_llama-100m-dpo-full
results: []
zephyr-smol_llama-100m-dpo-full
This model is a fine-tuned version of amazingvince/zephyr-smol_llama-100m-sft-full on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.5465
- Rewards/chosen: -0.0518
- Rewards/rejected: -0.7661
- Rewards/accuracies: 0.7170
- Rewards/margins: 0.7143
- Logps/rejected: -450.2018
- Logps/chosen: -588.7877
- Logits/rejected: -4.9602
- Logits/chosen: -5.2468
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: 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: linear
- 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 | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.6549 | 0.26 | 1000 | 0.6037 | -0.1205 | -0.4850 | 0.6550 | 0.3644 | -447.3903 | -589.4750 | -4.7410 | -5.0341 |
| 0.5349 | 0.52 | 2000 | 0.5779 | -0.0126 | -0.5080 | 0.6770 | 0.4955 | -447.6208 | -588.3951 | -4.8645 | -5.1463 |
| 0.6029 | 0.77 | 3000 | 0.5657 | 0.0902 | -0.4636 | 0.6900 | 0.5538 | -447.1767 | -587.3674 | -5.0016 | -5.2911 |
| 0.5273 | 1.03 | 4000 | 0.5596 | 0.0496 | -0.5449 | 0.7040 | 0.5944 | -447.9891 | -587.7738 | -4.9972 | -5.2892 |
| 0.5 | 1.29 | 5000 | 0.5557 | 0.0585 | -0.6110 | 0.7050 | 0.6695 | -448.6505 | -587.6843 | -5.0108 | -5.3047 |
| 0.5056 | 1.55 | 6000 | 0.5499 | 0.0054 | -0.6719 | 0.7130 | 0.6773 | -449.2598 | -588.2154 | -4.9988 | -5.2907 |
| 0.4608 | 1.81 | 7000 | 0.5500 | -0.0376 | -0.7494 | 0.7030 | 0.7118 | -450.0341 | -588.6455 | -5.0549 | -5.3406 |
| 0.426 | 2.07 | 8000 | 0.5472 | -0.0106 | -0.7021 | 0.7100 | 0.6916 | -449.5617 | -588.3751 | -4.9750 | -5.2626 |
| 0.3875 | 2.32 | 9000 | 0.5464 | -0.0011 | -0.7171 | 0.7140 | 0.7159 | -449.7113 | -588.2810 | -4.9935 | -5.2796 |
| 0.397 | 2.58 | 10000 | 0.5462 | -0.0391 | -0.7566 | 0.7190 | 0.7175 | -450.1064 | -588.6602 | -4.9737 | -5.2618 |
| 0.4486 | 2.84 | 11000 | 0.5459 | -0.0493 | -0.7667 | 0.7110 | 0.7174 | -450.2074 | -588.7629 | -4.9569 | -5.2441 |
Framework versions
- Transformers 4.35.0
- Pytorch 2.1.0
- Datasets 2.14.6
- Tokenizers 0.14.1
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 29.37 |
| AI2 Reasoning Challenge (25-Shot) | 25.00 |
| HellaSwag (10-Shot) | 28.54 |
| MMLU (5-Shot) | 25.18 |
| TruthfulQA (0-shot) | 45.75 |
| Winogrande (5-shot) | 51.07 |
| GSM8k (5-shot) | 0.68 |