Mistral-7B-DFT

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the siqi00/mistral_ultrafeedback_unhelpful_chatprompt_0.7_1.0_50_320 dataset. It was finetuned as part of the paper Discriminative Finetuning of Generative Large Language Models without Reward Models and Preference Data

The code is available at https://github.com/PenGuln/DFT.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-06
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2

Framework versions

• Transformers 4.45.2
• Pytorch 2.1.2+cu121
• Datasets 3.0.1
• Tokenizers 0.20.1

Usage Example

The model can be used for text generation tasks. A basic example using the transformers library is shown below:

from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
import torch

model_id = "siqi00/Mistral-7B-DFT"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True)

prompt = "What is the capital of France?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
generation_config = GenerationConfig(max_new_tokens=20, temperature=0.7)
outputs = model.generate(inputs["input_ids"], generation_config=generation_config)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)

Remember to install the necessary libraries (pip install transformers) and adjust parameters like temperature and max_new_tokens to fine-tune generation.

Citation

@misc{guo2025discriminativefinetuninggenerativelarge,
      title={Discriminative Finetuning of Generative Large Language Models without Reward Models and Preference Data}, 
      author={Siqi Guo and Ilgee Hong and Vicente Balmaseda and Tuo Zhao and Tianbao Yang},
      year={2025},
      eprint={2502.18679},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.18679}, 
}