metadata
license: apache-2.0
base_model: alignment-handbook/mistral-7b-sft-constitutional-ai
tags:
- alignment-handbook
- trl
- dpo
- generated_from_trainer
- trl
- dpo
- generated_from_trainer
datasets:
- HuggingFaceH4/ultrafeedback_binarized
- HuggingFaceH4/cai-conversation-harmless
model-index:
- name: mistral-7b-dpo-constitutional-ai
results: []
mistral-7b-dpo-constitutional-ai
This model is a fine-tuned version of alignment-handbook/mistral-7b-sft-constitutional-ai on the HuggingFaceH4/ultrafeedback_binarized and the HuggingFaceH4/cai-conversation-harmless datasets. It achieves the following results on the evaluation set:
- Loss: 0.6064
- Rewards/chosen: -6.2332
- Rewards/rejected: -9.9320
- Rewards/accuracies: 0.6825
- Rewards/margins: 3.6988
- Logps/rejected: -269.0856
- Logps/chosen: -253.0284
- Logits/rejected: -2.8561
- Logits/chosen: -2.8752
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: 2
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- total_train_batch_size: 16
- total_eval_batch_size: 64
- 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 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.0884 | 1.94 | 1000 | 0.5467 | -3.4828 | -6.0365 | 0.7025 | 2.5536 | -230.1307 | -225.5250 | -2.9599 | -2.9763 |
Framework versions
- Transformers 4.39.0.dev0
- Pytorch 2.1.0a0+32f93b1
- Datasets 2.14.6
- Tokenizers 0.15.2