AIPlans/Qwen3-0.6B-SFT-hs2

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Nov 29, 2025License:mitArchitecture:Transformer Open Weights Warm

AIPlans/Qwen3-0.6B-SFT-hs2 is a 0.8 billion parameter instruction-tuned causal language model, fine-tuned from Qwen/Qwen3-0.6B-Base by AIPlans. This model is specifically trained using selected high-quality responses to research model diffing, preference fine-tuning, and the evaluation of lightweight LLM behavior changes. It is primarily intended for research in understanding subtle modifications and their impact on smaller language models.

Loading preview...

Model Overview

AIPlans/Qwen3-0.6B-SFT-hs2 is a 0.8 billion parameter language model developed by AIPlans, fine-tuned from the Qwen/Qwen3-0.6B-Base architecture. This model was specifically created for the AI-Plans Model Diffing project, focusing on understanding and evaluating changes in lightweight LLMs.

Key Characteristics

  • Base Model: Fine-tuned from Qwen/Qwen3-0.6B-Base.
  • Training Method: Utilizes Supervised Fine-Tuning (SFT) with TRL, specifically trained on responses with a score of 3 or higher from its dataset.
  • Training Efficiency: Achieved training completion in approximately 1 hour and 10 minutes using an A100 (40 GB) GPU.

Intended Use Cases

This model is designed for research purposes, particularly in:

  • Model Diffing: Investigating differences between model versions.
  • Preference Fine-Tuning: Exploring methods for aligning models with specific preferences.
  • Behavioral Evaluation: Analyzing how minor changes impact the behavior of lightweight LLMs.