metadata
license: mit
tags:
- generated_from_trainer
base_model: HuggingFaceH4/mistral-7b-sft-beta
model-index:
- name: zephyr-7b-dpo-full-beta-0.2
results: []
zephyr-7b-dpo-full-beta-0.2
This model is a fine-tuned version of HuggingFaceH4/mistral-7b-sft-beta on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.7903
- Rewards/chosen: -3.2220
- Rewards/rejected: -7.3367
- Rewards/accuracies: 0.7659
- Rewards/margins: 4.1147
- Logps/rejected: -282.6258
- Logps/chosen: -314.5996
- Logits/rejected: -2.6943
- Logits/chosen: -2.6970
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: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- total_train_batch_size: 32
- total_eval_batch_size: 32
- 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.5631 | 0.26 | 500 | 0.5260 | 0.0288 | -1.2082 | 0.75 | 1.2371 | -251.9833 | -298.3453 | -2.9467 | -2.9577 |
| 0.5432 | 0.52 | 1000 | 0.5888 | -0.0335 | -1.8482 | 0.7540 | 1.8147 | -255.1831 | -298.6568 | -2.8465 | -2.8476 |
| 0.5368 | 0.77 | 1500 | 0.5860 | -0.4836 | -2.3300 | 0.7619 | 1.8464 | -257.5920 | -300.9073 | -2.8455 | -2.8445 |
| 0.0615 | 1.03 | 2000 | 0.6024 | -0.5971 | -2.6919 | 0.7778 | 2.0948 | -259.4018 | -301.4749 | -2.8687 | -2.8639 |
| 0.0817 | 1.29 | 2500 | 0.6655 | -1.3554 | -3.8426 | 0.7738 | 2.4872 | -265.1552 | -305.2667 | -2.8257 | -2.8254 |
| 0.0617 | 1.55 | 3000 | 0.6421 | -1.2552 | -3.7613 | 0.75 | 2.5062 | -264.7488 | -304.7651 | -2.7744 | -2.7683 |
| 0.0765 | 1.81 | 3500 | 0.6582 | -1.1492 | -4.0394 | 0.7659 | 2.8902 | -266.1391 | -304.2354 | -2.7403 | -2.7389 |
| 0.0178 | 2.07 | 4000 | 0.6797 | -1.8485 | -5.2549 | 0.7619 | 3.4064 | -272.2166 | -307.7317 | -2.7310 | -2.7273 |
| 0.0165 | 2.32 | 4500 | 0.7359 | -2.2096 | -6.0498 | 0.7817 | 3.8401 | -276.1910 | -309.5376 | -2.7006 | -2.7001 |
| 0.0094 | 2.58 | 5000 | 0.7864 | -2.8828 | -6.8542 | 0.7738 | 3.9713 | -280.2130 | -312.9036 | -2.7185 | -2.7196 |
| 0.0094 | 2.84 | 5500 | 0.7953 | -3.1897 | -7.3009 | 0.7579 | 4.1112 | -282.4464 | -314.4378 | -2.6987 | -2.7012 |
Framework versions
- Transformers 4.35.0
- Pytorch 2.1.0+cu118
- Datasets 2.14.6
- Tokenizers 0.14.1
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 61.55 |
| AI2 Reasoning Challenge (25-Shot) | 61.77 |
| HellaSwag (10-Shot) | 84.04 |
| MMLU (5-Shot) | 61.79 |
| TruthfulQA (0-shot) | 54.72 |
| Winogrande (5-shot) | 76.95 |
| GSM8k (5-shot) | 30.02 |