DeepSWE 1.1 trajectories: Qwen3.8-27B agents and baselines

This dataset contains agent trajectories and evaluation results from 7 complete runs on DeepSWE 1.1. The main experiments evaluate Qwen3.8-27B through Mini-SWE, Claude Code, and Pi. Muse-Glimmer-30B and Qwen3.6-27B are included as weaker reference baselines.

Every run covers all 113 benchmark tasks. Altogether, the dataset contains:

• 791 task-level result records;
• 791 compressed agent trajectories;
• 425 submitted text patches;
• exact run configurations, aggregate scores, and efficiency statistics.

This repository contains evaluation outputs. The tasks, environments, and verifiers live in the separate DeepSWE repository.

What is DeepSWE?

DeepSWE measures coding agents on 113 original, long-horizon software-engineering tasks drawn from active TypeScript, Go, Python, JavaScript, and Rust projects. For each task, an agent receives an isolated repository and a natural-language request. It must inspect the code, implement the requested behavior, run tests, and submit a patch. DeepSWE then applies that patch to a clean copy and grades it with held-out tests.

Results

An execution error means that the agent phase ended with an error. It is not the number of tasks receiving zero reward. Exact model, sampling, context-window, output-limit, timeout, retry, and agent-version settings are available through the linked run.json files.

Repository layout

README.md

summary/
    main-results.csv
    efficiency-quantiles.csv

runs/
    <run-id>/
        run.json
        tasks/
            <task-id>/
                result.json
                trajectory.json.gz
                model.patch          # when available

run.json contains the run configuration, aggregate scores, efficiency distributions, and an index of all tasks. result.json contains one task's scores, status, durations, usage metrics, and artifact availability. trajectory.json.gz contains the task prompt, assistant messages, reasoning trace, tool calls, tool outputs, and usage information. Message details vary slightly between agent interfaces.

model.patch is available for 425 tasks. Patch availability is recorded in each task's result.json.

Loading the data

Download a local snapshot:

from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="kaitchup/DeepSWE1.1-trajectories-Qwen3.8-27B",
    repo_type="dataset",
))

Read the result table and one trajectory:

import csv
import gzip
import json

with (root / "summary" / "main-results.csv").open() as file:
    runs = list(csv.DictReader(file))

run_dir = root / "runs" / "qwen3.8-27b-pi-xhigh"
task_dir = sorted((run_dir / "tasks").iterdir())[0]
result = json.loads((task_dir / "result.json").read_text())

with gzip.open(task_dir / "trajectory.json.gz", "rt", encoding="utf-8") as file:
    trajectory = json.load(file)

print(result["task_name"], result["verifier"])
print("trajectory steps:", len(trajectory["steps"]))

Limitations

• Each configuration has one run, so small differences should not be treated as statistically significant without repetitions.
• Agent interfaces differ in prompting, tools, context management, retry policy, and output limits. This is not a controlled model-only comparison.
• low, medium, and xhigh are configuration labels, not standardized amounts of inference compute across different agents or models.
• Token and turn counts come from different agent adapters and may not be perfectly comparable.

Acknowledgments

Verda provided the RTX Pro 6000s and H200s used to run these experiments.

Verda is a full-stack AI cloud built for high-performance inference, training, and agentic workloads, with data privacy and sustainability at its core.

License and third-party code

This dataset is released under Apache-2.0. The benchmark tasks use third-party open-source projects under their respective licenses. Trajectories and patches can contain excerpts from or modifications to those projects; those materials remain subject to their applicable upstream licenses.

See DeepSWE's project-level provenance table.