sagisch/Qwen3-4B-Dolci-Think-SFT-7B

Model Overview

This model, sagisch/Qwen3-4B-Dolci-Think-SFT-7B, is a 4 billion parameter language model derived from the Qwen/Qwen3-4B base model. It has undergone full-parameter supervised fine-tuning on a 600,000-example subset of the allenai/Dolci-Think-SFT-7B dataset. Its primary distinguishing feature is its ability to generate explicit reasoning steps enclosed within <think>...</think> blocks, mirroring the structure of its training data.

Key Capabilities and Features

• Reasoning Trace Generation: Designed to emit detailed reasoning processes within specific XML-like tags, which can be beneficial for interpretability and debugging.
• Extended Context Window: Inherits a substantial context length of 32,768 tokens from its base model, allowing for processing longer inputs and maintaining coherence over extended dialogues or documents.
• Fine-tuned for Reasoning: Optimized specifically for tasks that benefit from explicit thought processes, making it a strong candidate for research into chain-of-thought compression or similar reasoning-focused applications.

Training Details

The model was trained for a single epoch over 4,518 steps, processing approximately 6.0 billion tokens. Training utilized a cross-entropy objective, masking prompt and template tokens to focus on assistant responses. It was trained on 4 H100 GPUs with a global batch size of 128 sequences.

Usage Considerations

• Sampling Settings: For optimal performance and to avoid repetition collapse, it is recommended to use Qwen3's thinking-mode sampling settings: temperature=0.6, top_p=0.95, and top_k=20.
• Limitations: This model is a single-epoch SFT without preference tuning or RL. It has not been evaluated on standard reasoning benchmarks, and no safety tuning beyond the base model and dataset was performed. Its capabilities are largely inherited from Qwen3-4B and influenced by the Dolci-Think dataset's biases.