SWE-Eff†-14B is a LoRA fine-tuned SWE agent model based on Qwen3-14B, trained on ~3K high-quality filtered trajectories from R2EGym with a 32K context window. It uses suggestive thinking with masked supervision (mask_think) to inject reasoning prompts into training while masking them from loss, preserving the model's autonomous reasoning while implicitly guiding efficient agent behaviors.

SWE-Eff† serves as the conservative complementary model — optimized for harder problems involving multi-file logic, unclear root causes, and complex API interactions. For structured tasks with clear error traces, see the default model SWE-Eff.

Suggestive Thinking

Unlike the default SWE-Eff model, SWE-Eff† injects reasoning prompts (wrapped in <think...</think blocks) into each assistant message during training to encourage deeper reasoning:

Let's think step by step ... (encourages deeper reasoning before actions)
Let's view, think, edit, test ... (standardizes workflows)
If I get stuck in a loop, I need to think of different solutions to break out of it. (mitigates repetitive action loops)

The mask_think mechanism excludes the content inside thinking blocks from loss computation while retaining supervision on the <think and </think tokens. This allows suggestive guidance to influence the model implicitly while preserving autonomous reasoning behavior.

Training Data

Fine-tuned on filtered-R2EGym-SFT-Trajectories — 3,218 high-quality trajectories filtered from R2EGym-SFT via a multi-stage pipeline:

Basic Quality: exit_status = Submitted & resolved = True
Behavioral Soundness: Redundant loop detection & excessive search ratio filtering
Hallucination Control: Shortcut pattern & false reasoning detection
Thought–Action Alignment: Intent vs. action consistency enforcement

Training Configuration

Item	Value
Base Model	Qwen3-14B
Precision	bfloat16
PEFT Method	LoRA
LoRA Rank (r)	16
LoRA Alpha	32
LoRA Dropout	0.2
Target Modules	q/k/v/o/up/down/gate_proj
Adapter Size	246 MB
Global Batch Size	16
Gradient Accumulation	8
Learning Rate	2e-4
LR Scheduler	Cosine
Warmup Ratio	0.05
Weight Decay	0.1
Training Epochs	3
Total Training Time	~10.5 h
Hardware	2 × H200
Maximum Context Length	32,768 tokens
Key Modification	Suggestive thinking + `mask_think`

Evaluation

Evaluated on SWE-bench Verified using R2E-Gym scaffold with 32K context, 100-turn limit, temperature=0.6, top_p=0.95, and function calling disabled.

Metric	SWE-Eff (Default)	SWE-Eff† (Complementary)	SWE-Eff‡ (Union)
Resolved rate	21.6%	20.6%	30.4%
Avg steps	37.1	44.5 (+20%)	—
Submission success rate	43.2%	55.3%	—
Edit success rate	54.2%	63.2%	—
>80-step resolve rate	2.0% (1/51)	13.7% (7/51)	—

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")

model = PeftModel.from_pretrained(base_model, "ubicloud/SWE-Eff-Hard-14B")

When to Use

SWE-Eff† (this model): Multi-file logic, unclear root causes, complex API interactions, known hard projects (e.g., sympy, sphinx, psf)
SWE-Eff: Bugs with clear error traces, localized to a single file, structured repositories (e.g., django, scikit-learn, xarray)