fenyo/Qwen2.5-7B-base2instruct-SFT
fenyo/Qwen2.5-7B-base2instruct-SFT is a 7.6 billion parameter instruction-tuned language model developed by fenyo, based on Qwen/Qwen2.5-7B with a 32768 token context length. This intermediate checkpoint focuses on supervised fine-tuning (SFT) to teach the base model ChatML format and instruction following. It significantly improves instruction adherence, making it suitable as a foundational step for further instruction-tuned applications.
Loading preview...
Model Overview
This model, fenyo/Qwen2.5-7B-base2instruct-SFT, is an intermediate checkpoint in a larger base → instruct pipeline, specifically the Supervised Fine-Tuning (SFT) stage. Built upon Qwen/Qwen2.5-7B, its primary goal is to train the base model to understand and follow instructions using the ChatML format.
Key Capabilities & Training
- Instruction Following: The SFT process significantly enhances the model's ability to follow instructions, as evidenced by a substantial increase in IFEval scores from 27.4 (base) to 44.9.
- ChatML Format: It is fine-tuned to adhere to the ChatML conversation format, making it ready for chat-based applications.
- Training Details: The model underwent full fine-tuning using TRL's
SFTTrainerwithassistant_only_losson theallenai/tulu-3-sft-mixturedataset (180k examples) for one epoch.
Performance Highlights
While the SFT step dramatically improves instruction following (IFEval), it shows minor changes in GSM8K and MMLU scores compared to the base model. This checkpoint is a crucial step towards the final instruction-tuned model, which further refines performance through DPO and RLVR stages.
Usage
Users should apply tokenizer.apply_chat_template for proper interaction, following the ChatML format. This model serves as a strong foundation for developers looking to build highly instruction-following agents.