skymizer/Llama2-7b-sft-chat-custom-template-dpo

<!-- 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. -->

<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="200" height="32"/>

Llama2-7b-sft-chat-custom-template-dpo

This model is a fine-tuned version of elichen3051/llama2-7b-sft-chat-no-template on the HuggingFaceH4/ultrafeedbackbinarized, the HuggingFaceH4/orcadpo_pairs and the HuggingFaceH4/cai-conversation-harmless datasets. It achieves the following results on the evaluation set:

• —Loss: 0.4717
• —Rewards/chosen: -1.6807
• —Rewards/rejected: -3.1957
• —Rewards/accuracies: 0.6345
• —Rewards/margins: 1.5150
• —Logps/rejected: -519.5196
• —Logps/chosen: -379.2986
• —Logits/rejected: -2.7275
• —Logits/chosen: -2.7213

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
• —trainbatchsize: 8
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 7
• —gradientaccumulationsteps: 8
• —totaltrainbatch_size: 448
• —totalevalbatch_size: 56
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: cosine
• —lrschedulerwarmup_ratio: 0.03
• —num_epochs: 2

Training results

Framework versions

• —Transformers 4.42.0.dev0
• —Pytorch 2.3.1
• —Datasets 2.19.2
• —Tokenizers 0.19.1