Mistral-7B-v0.3-Middo-WizardLM

Paper: Middo: Model-Informed Dynamic Data Optimization for Enhanced LLM Fine-Tuning via Closed-Loop Learning

Code: https://github.com/Word2VecT/Middo

Model description

This model is a fine-tuned version of mistralai/Mistral-7B-v0.3 on the MiddOptimzed/mistral_wizard dataset.

Training and evaluation data

Training data

Middo optimized WizardLMTeam/WizardLM_evol_instruct_70k on mistralai/Mistral-7B-v0.3.

Evaluation data

• General
  • MMLU
  • IFEval
• Math
  • GSM8K
  • MATH
• Code
  • HumanEval
  • MBPP
• Reasoning
  • Hellaswag
  • GPQA

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 256
• total_eval_batch_size: 64
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 1.0

Training results

• epoch: 1.0
• total_flos: 4.871785990877872e+18
• train_loss: 0.6260631282554998
• train_runtime: 6928.3413
• train_samples_per_second: 12.871
• train_steps_per_second: 0.05

Framework versions

• Transformers 4.55.0
• Pytorch 2.6.0+cu124
• Datasets 3.6.0
• Tokenizers 0.21.1