JianhuiWei/qwen35_4b_sft_VE

Model Overview

JianhuiWei/qwen35_4b_sft_VE is a 4.5 billion parameter model derived from the Qwen/Qwen3.5-4B base, fine-tuned through a sophisticated three-stage Pareto curriculum Supervised Fine-Tuning (SFT) process. This specific checkpoint, from Stage 3 at global step 105, was trained using unweighted token cross-entropy loss and notably without Error Tool-Call Masking.

Training Methodology

The model's training data consists of distilled trajectories from Qwen3.5-122B, specifically those with a reward score of 0.3 or higher. The unique Pareto curriculum progressed through stages:

• Stage 1: Reward 0.3-0.5
• Stage 2: Case-wise Pareto 0.3-0.7
• Stage 3: Highest band per case 0.3-1.0

Custom Chat Template

A key feature is its custom chat_template.jinja, which differs from the stock Qwen3.5 template. This template is designed for both SFT rendering and inference, consistently replaying non-empty assistant reasoning_content within <think> blocks. It also supports XML-style Qwen tool-call formatting for parallel tool calls and groups tool responses into subsequent user turns. Thinking is enabled by default, but can be disabled by passing enable_thinking=False during prompt generation.

Use Cases

This model is particularly well-suited for applications requiring:

• Enhanced instruction following: Benefiting from the Pareto curriculum SFT.
• Structured conversational AI: Leveraging the custom chat template for improved tool-call handling and reasoning display.
• Integration with SGLang: The custom template ensures training/inference parity when served with SGLang, supporting qwen3 reasoning and qwen3_coder tool-call parsers.