nebius/Longtail-SWE-trajectories

Dataset Summary

Longtail-SWE Trajectories contains agent trajectories from model runs on the hard split of nebius/Longtail-SWE. Each record includes the agent's prompts, messages, tool calls and observations, together with its patch and saved evaluation result.

5,454 trajectories · 202 tasks · 9 model configurations · 3 runs per task

The runs use the common swelike scaffold with Claude Opus 5.5, GPT 5.6 Sol, GPT 5.6 Luna, GPT 6.1 Sol, Kimi K3, DeepSeek V4 Flash, GLM-5.3-Flash, Qwen 3.8 27B, and Qwen 3.6 27B. Configuration details are available in the row fields and observed_execution_settings.json; GPT 6.1 Sol settings are in sol61_addendum.json.

The results are saved evaluations from the task versions used in the runs, not regrades against every later task revision. Some initial statements differ from the exported task version; see initial_statement_matches_release. GPT 6.1 Sol has 3 skipped evaluations and 1 timeout, recorded as null resolved values. Details of its task-version differences are in sol61_addendum.json.

Resources

• Paper: Longtail-SWE
• Code and evaluation
• Harbor Hub datasets: hard · easy

How to use

from datasets import load_dataset
import json

trajectories = load_dataset("nebius/Longtail-SWE-trajectories", split="train")
row = trajectories[0]
print(row["instance_id"], row["model"], row["resolved"])

trajectory = json.loads(row["trajectory"])
for event in trajectory["events"]:
    print(event["type"])

Use instance_id to join trajectories with the hard split of nebius/Longtail-SWE.

Dataset Structure

Field name	Type	Description
trajectory_id	str	Unique trajectory identifier.
instance_id	str	Corresponding Longtail-SWE task identifier.
repo	str	GitHub repository owner and name.
programming_language	str	Task language.
task_month	str	Task creation month (YYYY-MM).
agent_config_id	str	Scaffold and model configuration identifier.
model	str	Model name recorded in the run configuration.
canonical_model	str	Normalized model name.
company	str	Model provider.
scaffold	str	Agent scaffold (swelike).
reasoning_effort	str	Explicitly recorded reasoning effort, or null.
context_limit	str	Recorded context limit.
paper_reasoning_label	str	Reasoning label used for the configuration; not necessarily the effective request setting.
effective_context_tokens	int	Recorded context limit in tokens.
run_index	int	Run number, from 0 to 2.
resolved	int	Saved result: 1 for resolved, 0 for not resolved, or null if unavailable.
exit_status	str	Agent exit status, when available.
source_cohort	str	Source run cohort.
raw_trajectory_sha256	str	SHA-256 of the original trajectory.
candidate_patch_sha256	str	SHA-256 of the original candidate patch before redaction.
initial_statement_matches_release	bool	Whether the saved initial statement matches the task release used for export.
trajectory	str	Normalized trajectory as a JSON string.
model_patch	str	Agent patch in unified diff format, after redaction.

Trajectory format (schema_version 1.2.0)

schema_version, trajectory_id, instance_id, language, repository
participant        {key, type, scaffold, model, reasoning_effort, ...}
run                {index, attempt, retry_number, source_cohort}
prompts            [{role, content, status}]
events             [{seq, type, ...}]
usage              {steps, model_calls, tool_calls, tokens, ...}
execution_settings {model, instance_context_limit, api_request_params, ...}
evaluation         {resolved, status, evaluation_matches_selected, ...}
provenance         {raw_trajectory_sha256, candidate_patch_sha256, ...}
normalization      {tool_result_limit_bytes, truncated_tool_results, ...}

The ordered events start with prompt and may include assistant_message, tool_call, tool_result, and final_message. Prompts and events are rendered from saved runs and redacted. Tool results over 64 KiB retain their first and last 32 KiB and are marked as truncated. Redaction can also change a model_patch, so a published patch is not guaranteed to apply unchanged.

License

The dataset compilation is licensed under CC BY 4.0. Source repositories retain their own licenses, and model and tool outputs may be subject to provider terms. See LEGAL.md.

Citation

@misc{dobrovolskii2026longtailswehighqualityswetasks,
    title={Longtail-SWE: High-Quality SWE Tasks for Underrepresented Languages},
    author={Vladimir Dobrovolskii and Ibragim Badertdinov and Alexander Golubev},
    year={2026},
    eprint={2611.10304},
    archivePrefix={arXiv},
    primaryClass={cs.SE},
    url={https://arxiv.org/abs/2611.10304}
}