6cf/Qwen3.5-4B-SWE-SFT-Coldstart

Qwen3.5-4B-SWE-SFT-Coldstart

Qwen3.5-4B-SWE-SFT-Coldstart is an agentic software-engineering cold-start checkpoint based on Qwen/Qwen3.5-4B.

The model was supervised fine-tuned on successful software-engineering agent trajectories generated by DeepSeek-V4-Flash and DeepSeek-V4-Pro. The trajectories use a Claude-Code-compatible tool harness and were executed in isolated SWE-Gym repository environments. Only trajectories accepted by an execution-based verifier were retained for supervised fine-tuning.

This repository contains the merged model weights. It is the SFT cold-start checkpoint before subsequent reinforcement learning.

Agent tool format

The checkpoint uses the native Qwen3.5 XML tool-call format. A typical call has the following shape:

<toolcall> <function=Read> <parameter=filepath>/testbed/package/module.py</parameter> </function> </tool_call>

Do not convert this checkpoint to a legacy Hermes JSON tool-call template at inference time. Use the included tokenizer and chat_template.jinja.

Training summary

• —Base model: Qwen/Qwen3.5-4B
• —Training method: LoRA supervised fine-tuning, merged into the base model
• —Teacher models: DeepSeek-V4-Flash, DeepSeek-V4-Pro
• —Domain: executable software-engineering agent trajectories
• —Environment: SWE-Gym repositories with a Claude-Code-compatible tool interface
• —Data filtering: execution-verifier-positive trajectories only
• —Training framework: MS-SWIFT
• —Precision: BF16
• —LoRA rank: 32
• —LoRA alpha: 64
• —Context length used for SFT: 32,768 tokens
• —Packing: disabled

Tool outputs and environment responses were retained as conversation context but were not intended as model-generated supervision. The supervised targets are assistant turns, including native Qwen3.5 tool calls and final responses.

Loading with Transformers

Use a recent Transformers release with Qwen3.5 support.

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "6cf/Qwen3.5-4B-SWE-SFT-Coldstart"

processor = AutoProcessor.frompretrained( modelid, trustremotecode=True, ) model = AutoModelForImageTextToText.frompretrained( modelid, torchdtype="auto", devicemap="auto", trustremotecode=True, )

For text-only agent inference, the same processor and model can be used with a text conversation and tool schemas passed through the included chat template.