AI Coding Dataset NEW

Covers two types of SFT and RL post-training signals: a track of correct answers, a trail of failure and attribution, supporting algorithms such as DPO/GRPO.

The training and evaluation are aligned with the assessment of the mission ' s profile lists, such as SWE-bench, DeepSWE, FrontierSWE, LMArena Coding.

From single file restoration to inter-module re-engineering hours-sky task, which is dedicated to strengthening context maintenance, planning and mid-trip error correction capabilities.

Docker + Verify Scripts for certainty, and the client can repeat the results of the assessment by one key to avoid environmental and calibration disputes.

Single-type missions, starting with 3,000-20,000, support the stratification and customization of the ease band (Easy/Medium/Hard/Frontier).

Python/Java/Go/ TypeScript/ Rust/C++ Six languages; warehouse-level tasks include complete scaffolding, unit testing and CI configuration.

The task level rubric+ step track + code diff 3 is marked, which is training, attribution, assessmentable.

The DeepSWE head alignment model 65-73% resolution rate ranges; the Ultra Long-horizon pairing Zimarason list is very different.

Complete process training for SFT/RL/Phosition to complete real engineering capability for the generic large model.

Capacity validation and short-board positioning for Claude Code, Codex, Opencode.

A high-division, pollution-resistant private assessment set was developed for the evaluation agency and model manufacturers.

To align the job set with the candidate model horizontally, output the selected conclusions that can be repeated.

JSON/JSONL structured sample + code repository snapshot (with git history) + Docker task mirror + check script.

task_id、prompt/issue、repo_snapshot、gold_patch、test_patch、trajectory、rubric、difficulty、metrics。

Double-reviewed note + automated script check + sample review, track replicability and label consistency are 98%.

Standard 2-4 weeks; warehouse level and hyperlong range 4-8 weeks; support for batch delivery and rolling expansion.

Commercial use authorization + compliance review of open source agreements + sensitivity to sensitive information and provision of complete authorization documents.

Other Organiser
"task id": "deepswe-00417",
"repo": "acme-org/serving-core",
"issue": "Multi-file connection: the leak in the connect pool led to long run."
"diff-git a/pol.py...",
"test pact": "tests/test pol leak.py",
"Tracory":  Chuckles 
"Step": 1, "action": "geep-R" Pool ("src/", "finding": "3 Unreleased"),
"Step": 2, "action": "edit src/pol.py:88", "diff": "..."
"Step": 3, "action": "pytest-q", "result": "28 passed"
I don't know.
"difficulty": "hard",
"Metrics": "resolved": true, "regression": 0, "Patch lines": 42)
♪ I'm sorry ♪

Other Organiser
"prompt": "Do a board application that supports multi-person collaboration, with drag and real-time synchronization,"
"candidate a": "stack": "React + WebSocket", "artifact": "..."
"candidate b":
"preference": "a",
"dimensions":
"Function integrity":
"UI Appearances":
"Interactive fluidity":
{\cHFFFFFF}{\cH00FFFF}
"Annotator agreement": 0.81
♪ I'm sorry ♪