This checkpoint is a full-parameter fine-tune of Qwen3.5-4B for autonomous repository-level issue resolution. It improves agent training data through trajectory-level quality filtering, step-level loss masking, and an offline curriculum learning.

Highlights

55.0% resolved on SWE-bench Verified (275/500).
+10.2 percentage points over Qwen3.5-4B and +4.0 points over the Raw RFT baseline.
3,093 retained trajectories after trajectory-level filtering.
1,250 low-quality assistant steps masked across 1,483 trajectories without deleting the surrounding interaction context.
Full-parameter SFT on 2 × NVIDIA A100 80GB GPUs.

SWE-bench Verified Results

All three model variants were evaluated on the SWE-bench Verified with the mini-SWE-agent harness.

Model	Resolved	Resolved rate
Qwen3.5-4B	224 / 500	44.8%
Raw RFT	255 / 500	51.0%
Refined SFT	275 / 500	55.0%

Trajectory Refinement

The training-data pipeline operates at two levels:

Trajectory-level selection. Teacher-generated trajectories are joined with execution-based evaluation results, retaining high-quality resolved demonstrations.
Action-level supervision. Clearly erroneous or inefficient actions receive zero training loss, while their observations and the agent's subsequent recovery remain in context.

This produces 3,093 retained demonstrations from 3,321 resolved candidates. A deterministic 90/10 split yields 2,784 training and 309 validation trajectories.

Offline curriculum learning

Training trajectories are ordered from shorter to longer sequences. Within each 64-sample length bin, examples are interleaved by rule-based coding action fingerprints, including direct patching, reproduction-first debugging, iterative debugging, exploration-heavy behavior, and linear fixes.

Training Details

Training Configuration

Setting	Value
Base checkpoint	Qwen/Qwen3.5-4B
Fine-tuning method	Full-parameter SFT
Framework	ms-swift
Hardware	2 × NVIDIA A100 80GB
Distributed training	DeepSpeed ZeRO-3, sequence parallel size 2
Precision	bfloat16
Training / validation examples	2,784 / 309
Epochs	2
Optimizer steps	1,392
Per-device batch size	1
Gradient accumulation	4
Maximum sequence length	65,536 tokens
Learning rate	1e-5
Schedule	Cosine decay, 5% warmup
Training runtime	Approximately 17 hours

Limitations

Results reflect one agent scaffold and evaluation configuration.
The comparison does not isolate every refinement component through separate ablations.

AxiaoDBL/qwen3.5-4b-swebench-sft

Qwen3.5-4B SWE-bench Trajectory SFT