RichardErkhov/HuggingFaceH4_-_mistral-7b-anthropic-8bits

Quantization made by Richard Erkhov.

Github

Discord

Request more models

mistral-7b-anthropic - bnb 8bits

• —Model creator: https://huggingface.co/HuggingFaceH4/
• —Original model: https://huggingface.co/HuggingFaceH4/mistral-7b-anthropic/

Original model description: --- license: apache-2.0 base_model: HuggingFaceH4/mistral-7b-cai tags:

• —alignment-handbook
• —generatedfromtrainer datasets:
• —HuggingFaceH4/ultrafeedbackbinarizedfixed
• —HuggingFaceH4/cai-conversation-harmless model-index:
• —name: mistral-7b-dpo-v21.0cai.0.2 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. -->

Mistral 7B Constitutional AI

This model is a DPO-aligned version of Mistral 7B on the HuggingFaceH4/ultrafeedbackbinarizedfixed and the HuggingFaceH4/cai-conversation-harmless datasets.

It achieves the following results on the evaluation set:

• —Loss: 0.6327
• —Rewards/chosen: -9.8716
• —Rewards/rejected: -14.5465
• —Rewards/accuracies: 0.6725
• —Rewards/margins: 4.6749
• —Logps/rejected: -329.8578
• —Logps/chosen: -294.6768
• —Logits/rejected: -2.1023
• —Logits/chosen: -2.1648

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: 2
• —evalbatchsize: 8
• —seed: 42
• —distributed_type: multi-GPU
• —num_devices: 8
• —totaltrainbatch_size: 16
• —totalevalbatch_size: 64
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: linear
• —lrschedulerwarmup_ratio: 0.1
• —num_epochs: 3

Training results

Framework versions

• —Transformers 4.36.2
• —Pytorch 2.1.2+cu121
• —Datasets 2.16.1
• —Tokenizers 0.15.0