SeanWang0027/Qwen3-4B-Base-NemotronIF-Reasoning-SFT-10k

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SeanWang0027/Qwen3-4B-Base-NemotronIF-Reasoning-SFT-10k is a 4 billion parameter instruction-following model based on Qwen3-4B-Base, fine-tuned for reasoning-on conversations. It was trained on 10,000 conversations from the Nemotron-Instruction-Following-Chat-v1 dataset, specifically designed to produce responses in a Qwen3 thinking format. This model excels at generating structured reasoning traces before providing an answer, making it suitable for tasks requiring explicit thought processes.

Loading preview...

Overview

This model, SeanWang0027/Qwen3-4B-Base-NemotronIF-Reasoning-SFT-10k, is a full-parameter fine-tune of the Qwen3-4B-Base model. It specializes in instruction-following conversations where explicit reasoning is required, utilizing a unique Qwen3 thinking format (<think>...reasoning...</think>\n\n...answer...<|im_end|>). The fine-tuning addressed a limitation in the base model's handling of special tokens like <think>, ensuring they are properly learned and emitted.

Key Capabilities

  • Structured Reasoning: Generates a detailed thought process before the final answer, adhering to the Qwen3 thinking format.
  • Instruction Following: Proficient in responding to instructions based on its training on the nvidia/Nemotron-Instruction-Following-Chat-v1 dataset.
  • Full Fine-tuning: Unlike LoRA, this model underwent full fine-tuning, allowing it to correctly learn and utilize special reasoning tokens that were initially shared and untrained in the base model.

Training Details

The model was trained on 10,000 reasoning-on conversations, resulting in 23,588 examples. Training involved FSDP2 over 2x A100-80GB GPUs, using AdamW optimizer with a cosine learning rate schedule. The training focused on completion-only loss for the reasoning trace and answer, achieving a final loss of approximately 0.6. The process took about 2.1 hours wall clock time for one epoch.

Good For

  • Applications requiring models to show their reasoning steps.
  • Instruction-following tasks where a structured thought process is beneficial.
  • Developers looking for a 4B parameter model with enhanced reasoning capabilities and proper handling of Qwen3's thinking format.

Finetunes
1 models