Hyeongwon/P2-split4_only_answer_Qwen3-4B-Base_0501-bs64-epoch6
Hyeongwon/P2-split4_only_answer_Qwen3-4B-Base_0501-bs64-epoch6 is a 4 billion parameter language model developed by Hyeongwon, fine-tuned from Qwen3-4B-Base. This model is specifically trained using Supervised Fine-Tuning (SFT) with TRL, focusing on generating direct answers. It is designed for tasks requiring concise, answer-only responses, leveraging a 32K context length.
Loading preview...
Overview
Hyeongwon/P2-split4_only_answer_Qwen3-4B-Base_0501-bs64-epoch6 is a 4 billion parameter language model, fine-tuned from the Hyeongwon/Qwen3-4B-Base architecture. This model was developed by Hyeongwon and trained using Supervised Fine-Tuning (SFT) with the TRL library, indicating an optimization for specific response generation rather than broad conversational abilities. It supports a substantial context length of 32,768 tokens.
Key Capabilities
- Answer-Only Generation: Specifically fine-tuned to provide direct answers, making it suitable for question-answering tasks where concise output is preferred.
- Base Model: Built upon Qwen3-4B-Base, inheriting its foundational language understanding capabilities.
- TRL Framework: Utilizes the TRL (Transformer Reinforcement Learning) framework for its training procedure, suggesting a focus on improving specific task performance.
Training Details
The model underwent Supervised Fine-Tuning (SFT). The training leveraged TRL version 0.25.1, Transformers 4.57.3, Pytorch 2.9.1, Datasets 3.6.0, and Tokenizers 0.22.2. Further details on the training run can be explored via the provided Weights & Biases link.
Good For
- Applications requiring direct and concise answers to prompts.
- Integration into systems where the model's output needs to be an isolated answer without additional conversational filler.