chlee10/T3Q-ko-solar-sft-v3.0
Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:10.7BQuant:FP8Ctx Length:4kLicense:apache-2.0Architecture:Transformer Open Weights Warm

chlee10/T3Q-ko-solar-sft-v3.0 is a supervised fine-tuned (SFT) language model developed by Chihoon Lee (chlee10) and T3Q. It is based on the chihoonlee10/T3Q-ko-solar-dpo-v3.0 model. This iteration focuses on refining performance through specific training hyperparameters, including a cutoff length of 4096 and a cosine learning rate scheduler. It is designed for tasks benefiting from its fine-tuned capabilities.

Loading preview...

Model Overview

The chlee10/T3Q-ko-solar-sft-v3.0 model is a supervised fine-tuned (SFT) version, building upon the chihoonlee10/T3Q-ko-solar-dpo-v3.0 base model. Developed by Chihoon Lee (chlee10) and T3Q, this model has undergone specific training to enhance its performance for particular applications.

Training Details

The fine-tuning process utilized a set of carefully selected hyperparameters to optimize the model's learning and generalization. Key training parameters include:

  • Batch Size: 16
  • Number of Epochs: 1
  • Cutoff Length: 4096 tokens, defining the maximum sequence length for training.
  • Learning Rate: 5e-5, managed by a 'cosine' scheduler with a warmup ratio of 0.06.
  • Optimizer: 'paged_adamw_32bit' for efficient memory usage.
  • LoRA Configuration: Employed with r=16, alpha=16, and dropout=0.05, targeting projection layers (k_proj, v_proj, gate_proj, down_proj, up_proj) to efficiently adapt the model.
  • NEFTune: Applied with a noise alpha of 5 to potentially improve robustness and generalization.

Intended Use

This model is suitable for applications requiring a language model that has been specifically refined through supervised fine-tuning, leveraging the architectural strengths of its base model and the detailed training regimen.