MistralForCausalLM_Cal_DPO

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset.

Model description

The Cal-DPO algorithm effectively addresses the alignment problem between large language models and human preferences by calibrating the implicit rewards in comparative preference learning to match the real rewards. It has demonstrated excellent performance in multiple task benchmark tests.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-07
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

We evaluate models on 6 key benchmarks using the Eleuther AI Language Model Evaluation Harness , a unified framework to test generative language models on a large number of different evaluation tasks.

• IFEval (https://arxiv.org/abs/2311.07911)
• BBH (Big Bench Hard) (https://arxiv.org/abs/2210.09261)
• GPQA (Graduate-Level Google-Proof Q&A Benchmark) (https://arxiv.org/abs/2311.12022)
• MuSR (Multistep Soft Reasoning) (https://arxiv.org/abs/2310.16049)
• MMLU-PRO (Massive Multitask Language Understanding - Professional) (https://arxiv.org/abs/2406.01574)

Framework versions

• Transformers 4.40.2
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.19.1