Lixiaojian123/qwen3-0.6b-sft

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 21, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Lixiaojian123/qwen3-0.6b-sft is a 0.8 billion parameter Qwen3-based causal language model, supervised fine-tuned (SFT) on selected 'no-think' instruction datasets from SmolTalk2. Developed by Lixiaojian123, this model is optimized for instruction following and chat-message interfaces, offering improved performance on tasks like ifeval compared to its base model. It is designed for applications requiring efficient instruction processing within a 32K context length.

Loading preview...

Model Overview

This model, Lixiaojian123/qwen3-0.6b-sft, is a supervised fine-tuned (SFT) version of the Qwen/Qwen3-0.6B-Base model. It was trained using TRL SFT on specific "no-think" splits from the HuggingFaceTB/smoltalk2 dataset, focusing on instruction following without generating extensive chain-of-thought reasoning. The model utilizes the standard Qwen3 Instruct chat template, enabling seamless integration with chat-message interfaces.

Key Training Details

  • Base Model: Qwen/Qwen3-0.6B-Base
  • Methodology: Full-parameter SFT using TRL / Transformers framework.
  • Dataset: HuggingFaceTB/smoltalk2 (SFT config), specifically using no_think splits like smoltalk_smollm3_smol_magpie_ultra_no_think and OpenHermes_2.5_no_think.
  • Context Length: Trained with a maximum length of 2048 tokens.
  • Steps: 5000 training steps with an effective batch size of 64.

Performance Insights

Evaluation against the official Qwen3-0.6B Instruct model shows notable improvements in certain areas for this SFT checkpoint:

  • ifeval: Achieved 39.19%, demonstrating enhanced instruction following capabilities.
  • truthfulqa_mc2: Scored 46.52%.
  • mmlu: Reached 49.35%.

Limitations

As a compact 0.6 billion parameter model, it may exhibit limitations such as hallucination, reasoning errors, and difficulty with complex instructions. Its training on "no-think" data means it is not designed to produce detailed step-by-step reasoning.

Finetunes
1 models